David graduated from MIT in Mechanical Engineering and Physics. He ran a product development firm for seven years, primarily developing IoT devices across a range of industries. He also served as a researcher at the Plasma Science and Fusion Center and as Director of Hardware at Kepler51. He’s now working on oroForge, a lightweight PLM built for startups and SMBs, with the goal of automating the tedious operational tasks of engineering so hardware startups can scale faster and with fewer mistakes.

⚙️ Mechanical Engineering Resources:

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it take to be The 10X Intern

The zero-to-many dilemma that hardware companies face

Zero to one. One to many. Everyone in startups can recite the phrases. In software, the second one is mostly a question of money and infrastructure: spin up more servers, hire more engineers, ship faster, eventually get a DevOps guy. In hardware, one to many is a different beast, and most founders don’t see it coming until they’re already bleeding.

This is not an article about capital. Yes, injection molds cost money. Yes, a production run requires cash you probably don’t have yet. That’s real, but it’s the obvious problem, and the obvious problem is rarely the one that kills you. The problem I want to talk about is operational and organizational, and it’s sitting inside your team right now whether you know it or not.

The transition from zero-to-one and one-to-many is hard because the work changes, the people who thrive in it change, and the discipline required changes. All three at once. That’s the whole article.

Where Zero to One Actually Ends

Zero to one does not end when your first prototype works. It doesn’t even end when customers say they like it. It ends at DVT, assuming you’ve met your customer requirements. EVT tells you the product works. DVT tells you it works for the people who are going to buy it. Those are different questions, and answering the second one takes longer than most founders budget for.

One to many is PVT. You’re no longer asking whether the product works. You’re figuring out how to manufacture it reliably at scale: teaching your contract manufacturer how to build it, managing suppliers, writing test requirements, running the supply chain week over week.

Those are not the same problems. And they do not reward the same skill set.

Simple Guide of the EVT, DVT, and PVT process

Death by a Thousand Paper Cuts

A colleague of mine described PVT better than I ever have: death by a thousand paper cuts.

When you cross into production, the task count explodes, and almost none of it is what you’d traditionally call engineering. You’re emailing your CM. You’re chasing your vendor. You’re scheduling with a certification lab (underestimate how long this will take at your own risk). You’re purchasing parts, pulling price estimates for your exec team, documenting features for the next launch. The list does not stop. Each task is small. Collectively they bury you.

I felt this most acutely at a previous company, where I was the main hardware guy. The work I was good at, the work that got us to a product, quietly stopped being the work that needed doing. None of it was hard in isolation. There was just an endless amount of it, and it was all operational.

That’s the part nobody warns you about. The difficulty of one to many isn’t intellectual. It’s logistical and relentless.

The Engineers Who Build Fast Are Not the Engineers Who Scale

The people who excel at EVT and DVT are tinkerers. That’s not a knock. Scraping a prototype together, moving fast, iterating, that is a genuine gift and it can be fun and engaging. But it is not the skill set PVT demands.

PVT is documentation-heavy. It’s managing the CM, managing suppliers, maintaining supply chain visibility, defining test requirements. It’s engineering operations. As a tinkerer myself, I’ll be honest: that work doesn’t excite me. For other engineers it’s their bread and butter. That’s the point. It’s not a question of competence; it’s a question of interest and specialty.

So when you’re hiring, there’s a question worth asking every engineer: do you see yourself staying on the technical path as a senior engineer, or moving toward engineering management? The engineering-manager profile, broadly, is the one that thrives in PVT. The senior engineer is usually the one who got you through EVT and DVT.

Here’s the catch: you need the tinkerer first. You can’t skip them. So most startups hire that engineer out of necessity, which is correct. The savvy founder hires the tinkerer knowing they’ll need a different engineer at PVT, and plans for it instead of being surprised by it.

Jiga

Source custom parts with direct access to the manufacturers making them. No middleman, no black box. Jiga gives you a vetted network and one accountable partner standing behind every order. You know who’s building your parts and communicate with them directly. AI-assisted operations handle supplier matching, quoting, quality verification, and production tracking, so the relationships stay human while the operations run at speed.

A Connector, a Turnkey CM, and a Lost Month

Here’s the kind of paper cut that draws blood.

We were in production with a turnkey CM handling all the purchasing. One of our parts was a specialized connector. Instead of sourcing it from the original vendor, the CM quietly bought it through Digikey. Convenient, until the Digikey stock dropped out. Now they had to go engage the vendor directly, mid-production, from a cold start. That added a month to our timeline.

The lesson wasn’t that the CM made a bad call, though they did. The lesson was that I should have ensured they were communicating with the vendor directly on a critical part from the beginning. That’s a PVT discipline. It has nothing to do with whether the product works and everything to do with whether you can keep building it.

While we’re here: most people forget about testing at the CM entirely. Defining the tests your CM runs to validate the product, often building a test rig to do it, is real PVT work that doesn’t exist in the zero-to-one world. If you haven’t thought about how your CM proves each unit is good, you’re not done.

The Wrong File

The most expensive paper cut I’ve experienced was also the simplest.

We were getting molds made for an enclosure. At some point we added alignment pins so the two halves of the case would seat cleanly during assembly. An engineer had two versions of the file sitting on his desktop. He sent the wrong one. The one without the pins.

It cost about two to three thousand dollars to retool. We got lucky: the CM could machine the pins into the existing mold rather than recut it. Otherwise it would have been ten to twelve thousand to redo the whole thing. Even the lucky outcome stings at startup scale, and the unlucky one eats the runway.

The failure wasn’t the engineer. It was that there was nothing stopping him. No formal ECO process. No PDM acting as the single source of truth. Two identical-looking files on a desktop and a fifty-fifty guess.

We were managing BOMs in spreadsheets. Our CAD lived in Onshape, which has a perfectly good version tree built in, and we weren’t using it seriously. The tools existed. The discipline didn’t. That gap is the whole story of why one too many breaks teams that crushed zero to one.

Why Nobody Teaches This

Software spent two decades telling everyone to move fast and break things. Lean startup added real value, especially the discipline of defining an MVP and validating it before overbuilding. But the famous line, “that if you’re not embarrassed by your first release you shipped too late,” does not survive contact with hardware. You can’t ship an embarrassing physical product to a few hundred customers and patch it over the air.

If you’re planning to produce more than a few hundred units, you have to document your process. Not because it’s satisfying, but because the alternative is the connector story and the wrong-file story on repeat.

EVT, DVT, and PVT are terms I never learned at MIT. I picked them up on the job, later than I should have. The startups I see doing version control and documentation well are almost always the ones who’ve shipped before and learned it the expensive way. That’s the entire reason I’m writing this: so the newer founders set up these processes before the first mold gets cut, not after.

The boring work, the documenting, the version control, the ECO process, is what separates an engineer from a tinkerer. Don’t learn the expensive way.

👶 Meme Of The Week

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it take to be The 10X Intern

👶 Meme Of The Week

🙋‍♂️ Interview Practice Question of the Week

Company: Relativity Space

What is the difference between MMC and LMC, and when would you use each?

✅ The Answer

Geometric dimensioning & tolerancing (GD&T) is another category of questions asked in technical interviews for design, production or manufacturing engineering positions with fair regularity, and as such we will be covering a few questions related to it over the coming months.

Most mechanical engineering students received an extraordinarily brief introduction to GD&T through coursework, and a lecturer probably spent ~1-3 hours covering all of drafting. If you were lucky, perhaps you received a homework assignment. If that sounds familiar, it was not adequate. Hobbyists can get away with FDM printing an .stl file, saying it warped a bit, and either reprinting or accepting the defect. Professionals cannot.

Any machined, cast, DMLS, or printed part is guaranteed to have a corresponding drawing provided to the manufacturer calling out critical features (if multiple manufacturing steps are required, an intermediate drawing is typically provided for before and after each operation). As a hardware engineer, being able to create and/or interpret these drawings is necessary for clear communication of what exactly is needed. Tight tolerances mean higher cost, but loose tolerances can result in non-functional parts (also increasing cost). It is our job to determine exactly how tightly a component needs to be controlled, and effectively communicate that. As such, one function of GD&T is to form a shared language between designers and manufacturers so that the functional aspects of a design can be focused on.

Build Real Supplier Relationships

Most hardware teams are stuck choosing between two bad options: instant-quote platforms that hide suppliers behind a black box, or local shops that hit capacity walls right when you need them most. One trades away relationships for speed. The other offers trust but can’t scale.

Jiga gives you both. Direct access to vetted manufacturers with reliable capacity from prototype through production. You speak directly to the people making your parts. And when you come back, they already know your specs and technical needs. Same shop, same knowledge, consistently reliable quality without starting over every time.

Try Jiga Today!

As you can imagine, one of the most common design features is the hole. The maximum material condition (MMC), is extremely aptly named, and refers to the condition at the end of the tolerance zone where the most material remains on the part (think of it as the condition where the part has the most volume or mass). For a hole, this would be when it is the smallest, such that the remaining component is as large as possible. The opposite case is the least material condition (LMC). This is when the remaining component is as small as possible, and usually refers to when a hole is the largest it can be while remaining within its tolerance band.

In most cases, some form of pin, shaft, rod or bolt goes through the aforementioned hole. Given that assumption, if I’ve designed a hole to be between 9 and 10 mm, the larger hole must have some form of “slop” or “looseness” since a 9 mm hole is also functional. As such, a 10 mm hole can have a looser positional tolerance given its larger size suggests unnecessary slop in the system. This looser positional tolerance is referred to as a “bonus tolerance,” and is applied when an appropriately called-out feature departs from MMC.

What that means is that if the previous hole is larger than MMC by 0.5 mm (for a total diameter of 9.5 mm), assuming a positional tolerance at MMC of 0.5 mm, the total positional tolerance of that specific part becomes 1 mm (0.5 mm at MMC + 0.5 mm bonus tolerance due to departure from MMC). Use of MMC in this fashion allows us to accept parts using dimensioning based on functionality rather than regular dimensions, and leads to more parts accepted / lower production cost.

MMC’s far less common cousin, LMC, is typically used in situations where an interference fit is desired or where a hole lies close to the edge of a component. In this case, as the hole gets smaller, its positional tolerance loosens to allow the hole closer to the edge without concern of wall thickness reducing to dangerous levels.

In summation, MMC refers to the maximum material condition of a part, when it occupies as much volume as possible while remaining in specification. It is most often used to guarantee fitment when some form of rod or shaft goes through a hole. LMC refers to the least material condition of a part, when it occupies the least possible volume while remaining in specification. It is most often used to guarantee an interference fit or to prevent damage/encroachment onto a thin wall when a hole is placed close to the edge of a part. In both cases, as hole size departs from the material condition, a “bonus tolerance” is added to the normal tolerance, effectively loosening it.

This week’s article is a special edition written by Daniel Lieben about test engineering. Daniel completed his BSME at Northeastern University in 2023, doing two internships in his time there including one at Tesla. That internship led to a full-time role with Tesla as a test engineer, focused on their vehicle heat pump systems.

⚙️ Mechanical Engineering Resources:

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

Why Should Hardware Teams Run Tests?

This may seem like a silly question with an obvious answer, but it’s far from it. The easy answer is “so we know the product works,” but in practice, “works” is both very vague and a very low bar. Does it work for the entire intended lifespan? Does it work in all environmental conditions? Does it work when the user abuses it? These are all better questions to ask than simply “does it work?” And luckily - tests can answer all of these questions if executed well.

While there are many factors that make a useful test, there are two golden rules that I believe are non-negotiables:

1. Tests should answer questions. A test is useless if it doesn’t yield new information. Questions should generally be as specific as possible. For example, if you wanted to test the range of an electric vehicle, here are some questions you could ask, from least useful to most useful:

   1. I don’t have a question, but my boss says this test needs to be run (0/10, you don’t even know what you’re looking for)

   2. How far can this car go on one charge? (3/10, at least it’s a question, but it’s very vague)

   3. How far can this car go on one charge on a flat road at 70 mph on a 20°C day? (6/10, good specificity, but how useful is the result?)

   4. What is the car’s range sensitivity to road grade between -5 and 5%, vehicle speed between 40 and 80 mph, and ambient temperature between -20 and 30°C? What’s the best-case and worst-case range within those constraints? (10/10, now that’s a question that provides useful information)

2. Tests should be actionable. This one is a pet peeve of mine. If your project’s next steps are the same regardless of a test’s outcome, then it’s a useless test. Ideally, before starting a test you should know what all the possible outcomes lead to, and what defines each of those outcomes (i.e. your pass/fail criteria). I’ve watched engineers spend weeks setting up and running tests, only to be told that the project is proceeding as-is despite their subpar result - fantastic way to waste time and money.

How to Run a Test

Now that we know what makes a good test, how do we actually run one? Testing is extremely broad and not easily standardizable, but almost all tests will have a few components in common:

1. Your device under test (DUT). This is the part/system that is actually being observed; the thing that you have a question about. In the range test I used as an example before, this would be the car itself.

2. Instrumentation. Instrumentation has two buckets, sensors and data acquisition (DAQs). If you need to measure forces, temperatures, pressures, or whatever else, you’ll need to pick appropriate sensors for the job. DAQs are devices to read the sensor’s raw output - usually a small voltage or current, but there are more complicated types - and translate it into data that can be read/saved on your computer. Instrumentation is a very deep topic that I don’t have space to do justice in this article, but people make whole careers out of it. In the range test example, this would be an inclinometer to measure road grade, a speedometer to measure vehicle speed, and a thermocouple to measure ambient temperature, plus DAQs to read them all. For shock and vibe do NOT underestimate the importance of sampling frequency.

3. Computer & test script/software. Basically all tests will involve using a computer to do some of the following: log data, control the DUT & its environment, detect test failures/issues, track test progress, apply safety controls, visualize data, and more. Sometimes there are off-the-shelf softwares that can fit your needs, but often it’s necessary to write your own code to do exactly what your test calls for. Python is fantastic for test scripts, as there are tons of pre-existing libraries you can use to communicate with any given device in your test setup. In the range test example, you’d probably want to write a script to log all your sensor data over time, plus some values like battery percentage from the car.

Integrating these pieces together takes some persistence when you’re doing it for the first time, but gets quite a lot easier each time you go through the process. There is also a balance to be found between automating 100% of a test versus keeping a human involved in it, which brings me to my next point.

Test Automation

Let’s say you want to run a test to answer the question “will my part survive its 20-year lifespan with no degradation to its performance?” The test to answer that question will take quite a long time to run - not 20 years, but at least a few months (fun fact: I have a test that has been going for two years!). It would obviously suck to manually operate a test for that long, so this is a case where fully automating the test makes sense. How does one actually do that?

The most important thing that changes between a manual test and an automated test is safety. An automated test will continue to run even when you’re not sitting there babysitting it, so it needs to be safe for passersby in the lab as well as the lab itself. Automatic failure detection is a must - if anything goes wrong with the test, the test script needs to shut down the DUT safely and dissipate any hazardous energy, whether that energy is electrical, mechanical, thermal, or other.

Aside from safety, things like test tracking and data logging also become critical. If something interesting happens in the test and you’re not there to watch it, you need to be able to dig into the data and see exactly what happened and when.

Late parts don’t have to be a surprise

The worst part about a missed deadline isn’t the delay, it’s finding out too late to do anything about it. Black-box sourcing platforms give you a lead time estimate and then silence. On Jiga, you communicate directly with the manufacturer and get proactive updates on production status. When something shifts, you know early enough to adjust, not after your team is already scrambling.

Try Jiga Today!

Why Test and Simulation are Besties

Although test and simulation seem like two unrelated subdomains of engineering, they actually go together very well. Simulation serves a similar purpose to testing: to answer questions. The methodology is different, but the goals are the same. As for how they can work together, the short version is: tests validate simulations, and simulations extend tests.

For example, let’s imagine you want to figure out how hot the inside of a car will get when it’s parked on a sunny day. This is a relatively easy test to run - grab your car, put some thermocouples in it, park it in the sun, and let it sit for a few hours. Simulating it, on the other hand, is pretty complicated. A car is a huge system with lots of thermally coupled parts, and you’d need to model all three modes of heat transfer - yuck. To make an accurate model of that without any test data to compare to would be a huge challenge for even the best simulation engineers. With test data though, it’s a different story. It’s much easier to make a low-fidelity model that you can tune to match the test data you’ve collected, and then you can be confident that your model is accurate.

Next, let’s imagine you want to figure out how hot that same car would get if all the windows had 5% higher reflectivity. This is much more of a pain to test; you’d need to get new custom-made windows and install them on the car. However, if you have a model that you know is accurate from the previous test, you can just change the windows’ material properties and rerun the simulation! In this way, the simulation has extended the test. Note that the further you extrapolate the model from the test data, the less accurate the model is likely to become.

Why Testing is Awesome for Early-Career Engineers

Many people say that test engineering is a very good early-career job for a mechE, and I wholeheartedly agree. Test engineering is inherently very generalist; in my 3 years at Tesla I’ve worked with mechanical hardware, electrical hardware, firmware, reliability, failure analysis, software, simulation, and more. It’s a nice blend of hands-on lab work and computer work, and there are usually travel opportunities for tests which require specialized facilities. Even if your long-term aspirations are in design, test engineering is a great early-career stop to teach you about what makes a design good, though some engineers have reported difficulty switching between disciplines later.

Thanks for reading, shoot me an email at hardwareishard@gmail.com if you have any questions or follow up content about test engineering you want to see!

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

Meet our sponsor, Jiga! They connect you directly with manufacturers. Real quotes, from real humans, for real hardware, fast.

Try Jiga Today!

🦸 Guest Writer: Madhav!

This week’s newsletter is a special edition written by Madhav Bhat based on an interview he had with Pari Singh, the CEO of Flow Engineering (as well as several other AI CAD tool and adjacent interviews he’s conducted)! Madhav graduated from the University of Michigan with a BSE and MSE May of 2024. Throughout his recruiting journey he interviewed for mechanical engineering internships and full-time positions with companies from large Fortune 500 corporations (think Ford, Michelin, Apple) to growth-stage companies (think Tesla, SpaceX, Anduril) to startups ranging from 3-1000 people in size working on all sorts of projects.

Introduction

After our article on what do AI tools look like for mechanical engineers, we got a lot of interest in “software for hardware” and demand for hardware related AI tools. I sat down recently with Pari Singh, founder and CEO of Flow Engineering as well as the Flow Engineering team, to talk about where hardware development is headed. Flow is a requirements management and systems engineering platform that’s been quietly showing up at a lot of the companies you’d expect - Skydio, Rivian, Joby - and a lot more you wouldn’t. Pari’s a Forbes 30 Under 30, recently raised from Sequoia, and has some pretty hot takes on how hardware engineering is going to change. This piece is half interview takeaways, half my own thoughts on what it means for the rest of us.

Every MechE who has spent five minutes in a serious hardware program knows the cadence. Six months of requirements review, a frozen ICD, a design that nobody is allowed to touch, a CAD release, then the inevitable scramble when something downstream breaks the assumption upstream. That’s waterfall, and it’s how NASA built rockets in 3 years. SpaceX builds them in 3 months.

The reason isn’t (just) that SpaceX moves fast and breaks things. It’s that they swapped processes. Pari frames it as a three-stage evolution:

• Waterfall: top-down, sequential, “design the CAD file once and never touch it.” NASA on Apollo. 1 revision is the goal.

• Agile: parallel, iterative, requirements that move daily. SpaceX, Anduril, the modern deep tech stack. Maybe 5 revisions per design.

• Agentic: swarms of AI agents continuously redesigning, re-simulating, re-releasing. 500 revisions. The half-life of code (and CAD) drops from years to days.

Pari’s number for software is a useful reference point: at Google, 100% of code is refactored or rewritten every 5 years. In an agentic world, he claims that drops to 5 days. Whether or not the exact number holds, the direction is the part you should care about.

“Design iteration every 3 months becomes design iteration every 3 days.”

This has already happened in software, in stages. At OpenAI, GPT-3 wrote nothing on its own, humans drove. GPT-4 more AI started writing the software. By GPT-5 it was all agents. The interesting question is when this hits hardware, and how. I thought this chart from their AI Systems Engineering handbook was a good breakdown.

What is Flow?

Flow originally didn’t start as a software company and was born from Pari’s rocket design consultancy in London. They had built internal tools to design rocket engines 6x faster than the industry standard and quickly realized that these internal tools that would become Flow were the real game changers. Flow is, on the surface, a requirements management tool. So is DOORS. So is JAMA. So is Polarion. The legacy players in this space are decades old and their UIs feel decades old. Flow’s customer list (Skydio, Rivian, Joby, Astranis, and a bunch of others) skews young, fast-moving, and tech-forward. The pitch from Pari is not “we’re a better DOORS.” When I asked him directly whether he wanted people to think of Flow as the modern alternative to those tools, he was pretty firm: don’t anchor on the old solutions at all.

What he wants is for you to think of Flow as a systems graph: a live, queryable, single source of truth about how your product is structured, what it has to do, and how each requirement traces down through subsystems, components, tests, and analyses. CAD lives in your CAD tool. Sim lives in your sim tool. Test data lives in your test tool. But the system, the graph that connects all of it, lives in Flow. And because it’s a graph and not a spreadsheet, agents can crawl it, reason over it, and act on it. A good analogy that Pari mentioned was “when you ask Waymo to send a system, what would they send you?” They’re not going to send CAD, or a wiring diagram. This is what Flow is describing.

The pain point this addresses is: how does one change affect all the downstream dependencies? Payload mass on a satellite goes up by 4kg three months in. Who needs to know? The structures team, the GNC team, the propulsion team, etc. In a typical company today, that information propagates through so many communication channels or is not surfaced in the right channels and dependencies may not get ironed out. Flow’s bet is that an agent should propagate it instead, flag the affected requirements, kick off the affected analyses, and surface the trades.

👶 Meme Of The Week

Agentic Engineering

One of the big takeaways I wanted to get from Pari is what does the future of engineering work look like and it kept going back to this phrase “agentic engineering”. I already talked about it a bit above (having agents do all the work and reduce the hardware development cycle time to a fraction of what it currently is) but how does that work in practice? Flow is focused on solving the design side first, where the output is a CAD file. Whether that means partnering with companies working on CAD or building it themselves, that’s up in the air.

One of the largest pain points that Flow has seen at customers is siloed and disconnected teams. The bet is that agents handle the propagation and flagging, freeing engineers to make actual decisions. Right now Flow released a handbook on AI Systems Engineering and describes that there are 5 levels to AI Systems Engineering (which honestly reminded me of the 5-SAE levels of autonomy, The parallel doesn’t seem accidental, both are frameworks where each level only works if the one beneath it is solid.).

Level 0 — Manual. No AI in the workflow. Spreadsheets, manual traceability, periodic reviews. This is the baseline most legacy programs still operate at. Iteration time: months.

Level 1 — AI-Assisted. Engineers use ChatGPT or Claude on the side. Productive-feeling, but nothing structural changes, no shared standards, no record AI was involved. This is where most of the industry sits right now. Iteration time: still months.

Level 2 — AI-Native. No artifact gets authored from scratch by a human. Every requirement and test case starts as an AI output that an engineer refines. Slower per-task than editing a cell directly. The payoff is a complete record of AI output vs. human edits, which is the training data everything above this level needs. Iteration time: months → weeks.

Level 3 — Agentic. Engineering standards get codified into reusable skills — your shall-statement format, naming conventions, DO-178C clauses, etc., encoded once so AI applies them every time without being re-prompted. Iteration time: weeks.

Level 4 — Autonomous. Workflows stop requiring human initiation. A Git PR fires a requirements check. A CAD update triggers a mass-budget verification. Checks run on 100% of changes, not the 10% someone remembered to verify. Fair warning: teams hit this level and initially see more issues, not fewer since the system is catching things that were previously slipping through. Iteration time: days.

Level 5 — Harness Engineering. The AI system that builds the product becomes the product. A new role emerges — the harness engineer — who tunes the meta-system: false positive rate, mean time to detection, context coverage. They don’t write requirements; they engineer the team of agents that does. This is the current frontier. Iteration time: days to hours.

This has a lot of parallels to the way that software engineering is done nowadays. Half the time if you’re using Claude Code you’re not reviewing every line that it’s writing. The goal is to go to that with hardware engineering where everything is agentic. You tell it what you want to do and it does all the processes and you just review the output. The goal of the software engineer now is as an architect, and soon the hardware engineer as well.

It’s also very interesting to think about how this is going to extend to the manufacturing world. A world where these agents take validated designs, send them to JIGA or another manufacturing partner (or even find and negotiate with a contract manufacturer on your behalf), and have a physical product as the output is

shockingly

close. What I don’t see just yet is how this translates to scaled manufacturing (100s to 1000s in volume). But that is truly only a matter of time.

Forward Deployed Systems Engineers

A small thing that I want to flag because I’d never seen it before: Flow has a role called Forward Deployed Systems Engineer. If you’ve been around Palantir folks you’ll recognize the pattern (FDEs there are veeeeryy well known), but this is the first time I’ve seen this specific role. The idea is that your systems engineers go embed at the customer, learn their actual engineering problem, and shape the product around it. I’ve seen this kind of engineering role become super popular (shoutout Applied Intuition) as it provides a diverse skillset of deep technical product knowledge and customer-facing chops. That’s a pretty interesting career path if you’re a MechE who likes the consulting-meets-product flavor of work, and I’d keep an eye on it as a job category that’s about to grow. If you’re interested in a role like this feel free to reach out!

Integration and Customers

What does integration of an AI tool like this look like? To be honest it’s not too difficult. It’s an API. Customers can own the integration if they want but it’s an API that connects to all your different data sources like MATLAB scripts, CFD results, Python scripts, etc. It then synthesizes these results and tracks them in the platform. Cursor can do the integration in a couple minutes. The concept is called traceability and can briefly be described as being able to trace high-level requirements down to the individual simulation or test results that lead to those requirements being fulfilled.

For example: if my requirement is that my spacecraft needs to withstand 4G during launch, I want to link that with a launch simulation or actual results from a test rig. Being able to link multimodal data into the platform becomes pretty powerful.

Flow is used by some pretty huge companies like Astranis, Joby, etc. but almost all are super tech-forward and known to be faster-paced environments. When I asked about how you penetrate into legacy companies using legacy workflows, the answer was “we’re not really trying to do that”. Majority of Flow’s sales are coming from inbound and they are only partnering with companies that have an explicit desire to move much faster. However, Pari did mention that the legacy market is starting to wake up and the tides are changing day-by-day.

What’s Next?

Pari describes Flow as a movement to change how Systems Engineering looks and their mission is to help engineering teams move faster. When I asked what’s next for Flow, he said the goal is to go from end-to-end autonomously by the end of the year. Requirements to simulation. Even more high-level, the goal is to give Flow to a 14-year old and have them design a rocket or humanoid robot and essentially “vibe engineer” it.

In terms of the biggest lesson when building Flow and what is the best part of building Flow, Pari says “We are genuinely changing the way engineering is done. If you imagine the Starship Enterprise from Star Trek, we as a species can’t do that yet. It’s clear that Excel -> CAD etc. is dead. There’s a new fundamental approach and change in how engineering is done for the most important companies in the world. We get to work with our heroes”.

Flow just released an AI Systems Engineering handbook that you can check out here and encourages any Student Project Teams that are interested in using Flow to reach out!

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

Technical Interview Question

Context

Redesign this flawed 2x4 Lego brick with manufacturability in mind!

Pictured below is a poorly designed 2x4 LEGO brick. As the last Lego engineer standing, you hold the key to unlocking its true potential. With manufacturability, specifically injection molding, as your guiding star, you are entrusted with the noble task of reshaping this humble brick. Let innovation ignite your imagination as you breathe life into a new era of Lego brilliance.

Main Question

Pictured above is a poorly designed 2x4 LEGO brick. Improve its design with manufacturability (injection molding) in mind.

Answer

We could give you the ~quick and dirty answer to this one, but we think it’s an awesome opportunity to dive into injection molding and a little LEGO history (or lore, if you will).

Injection molding is an important concept/skill that is rarely discussed in school, but used frequently in MechE jobs, especially in product design. This manufacturing method is so prevalent primarily due to its ability to produce high-precision parts, in large quantities. For lower quantities, 3D-printed or machined plastic (depending on how representative the material properties need to be) can make more sense as creating a mold can be expensive, and only pays off if that cost can be amortized across thousands of parts.

As a design engineer, technically, you are mainly responsible for designing the part (manufacturing/process/tooling engineers are more involved). That being said, there are some fundamentals that you need to know. The mold consists of two large blocks of metal: the core, and the cavity. These are usually CNC’ed out of aluminum (soft tooling) or steel (hard tooling) and can then be hardened via heat treatment after machining. If hardened, modification to a part after the tooling is made can be extra expensive because the material being re-machined is now hardened steel (it’s still cheaper than getting a new mold). This category of alterations is referred to as “tool-safe” changes, and tolerances should be designed accordingly as more material can always be taken off, but adding is difficult). The plastic that is getting molded starts off in pellet form and is then heated and pushed through a large heated screw via a hydraulic piston. High pressure is essential to make sure the plastic fills the entire mold. Once the mold is filled, the plastic is cooled and then ejected from the mold, often with pins or high-pressure air.

Late parts don’t have to be a surprise

The worst part about a missed deadline isn’t the delay, it’s finding out too late to do anything about it. Black-box sourcing platforms give you a lead time estimate and then silence. On Jiga, you communicate directly with the manufacturer and get proactive updates on production status. When something shifts, you know early enough to adjust, not after your team is already scrambling.

Try Jiga Today!

Now that you’re familiar with the basics, back to LEGOs!

LEGOs are considered to be some of the best-designed injection-molded projects in the world. The average maximum force the bricks can stand is 4,240N (950 lbs). If you divide that by the mass of a single brick (1.152g), then (without accounting for buckling), you could stack 375,000 bricks high before yielding occurs. Also of note, a fan tested how many times a LEGO brick could be assembled before failing. The answer: 37,112 times!

So now we know, that you know, that LEGO design is sick! We hope you’re as excited as we are. Let’s dive into the details!

The wall thickness of a LEGO brick is 1.2mm and is uniform throughout the part (depending on the model/release date). This is essential so that plastic flows uniformly through the part, and upon cooling, shrinks at the same rate. Parts are filleted where appropriate (outer radius = inner radius + material thickness) to maintain even wall thickness throughout. This reduces stress concentrations, improves flow, and makes ejection easier.

All walls of the part in the tooling direction are drafted by 0.5 degrees so that it can separate after cooling. This is a very small draft angle (typically use 3-5 degrees). But, LEGO parts are also small, and the material (ABS) is very uniform. Larger draft angles are required for deeper parts or when the surface finish isn’t smooth (more difficult to eject without damaging the part in those cases).

Another necessary component of enabling separation is the parting line! You must design this into your parts from the start, or it can require significant rework. The parting line is where the core and the cavity meet. It often leaves a small and noticeable flash line, so place it accordingly. In LEGO’s case it is well hidden, and occurs on the bottom edge (shown in picture B), just before any radius shows up on the part.

Ribs are essential to increasing part strength when injection molding. Here are some basic rules of thumb:

• Ribs should be 40-60% of the wall thickness to which they connect.

  • Often, you’d want a thicker rib to add rigidity, but if it’s too thick, a sink mark will appear on the wall to which it is connected.

• Ribs should not be more than ~three times the thickness of the wall they attach to in height. This ensures they are fully filled during the injection. Just like any other wall, don’t forget that ribs need to be drafted too!

  • This defect is referred to as a “short shot,” namely because the injection did not make it far enough before cooling to fill the entirety of the mold (it came up short). Depending on part size, plastic viscosity/injection temperature, and injection pressure, beware of designing features with thickness < 1mm. Also, pay attention to the proximity to gates (more on that later)

Round cylindrical protrusions jutting from a part are called “bosses” (shown in picture A) They are typically used either as a receptacle for a screw/insert, or to help locate/assemble the part. The latter is the case for a LEGO. Bosses, like ribs, should be 40-60% of the wall thickness to which they connect. When it comes to draft, both the inside and outside of the boss need to be drafted. Bosses are often supported with ribs, as is the case in this part.

Gates are the injection point of the plastic. Depending on which release date of LEGO brick we look at, the gate has moved around a bit. In the part shown in picture C, it is right in the middle of the brick and can be noted as a small nub where the plastic is broken to separate the piece. In larger bricks, LEGO is able to hide the gate more effectively, but in the infamous 1x1, there is extremely limited real estate.

Now that the core and cavity have released the part, it is likely still attached. To push it off of the mold (and into a bucket presumably), ejection pins are the go-to method. These are small cylindrical pins that push out of the mold and knock the plastic off the mold. They leave small marks as the plastic has not cooled fully yet, and can be found on pretty much any injection molded part. In LEGO’s case, it may also be a sleeve around the central stud as can be seen by the slight marking in picture A.

Defects/terms that you should know of and be able to offer a brief explanation & fix: flow lines, weld lines, sink, short shot, flash, and jetting. This is a very common interview question if injection molding part design is involved.

Ok, so now to answer the actual interview question we started off this whole newsletter with:

I like to start off by running through the basics with uniform wall thickness, fillets, and draft. Then, jump into ribs and bosses to increase strength/function. Finally, the place to impress is by getting creative with the parting lines, gates, and ejection pins. I usually like to bait the interviewer to ask about defects and have that be my go-to follow-up!

If you really want to get fancy, you can jump into the material properties of different plastics, discuss other strengthening methods like glass filling or even bring up insert molding/over-molding if applicable to the part (not as applicable in this question but often used in the industry).

Follow Up(s)

It is also worth noting, that there are two important steps involved in kicking off a part with suppliers that we have thus far only alluded to:

• Design for Manufacturing (DFM)

  • After sending out a design, a good vendor will send back a presentation detailing any requested changes in design and tolerances (from your accompanying GD&T drawing). You can then push back, or accept changes until the final design is agreed upon. In this stage, you will likely see photos of their proposed mold geometry.

  • Never fully trust a vendor. At the end of the day, it’s your job to make sure your part works. Depending on the vendor, they may agree to all your tolerances and then ship you parts completely out of spec. Combat this by requesting first article inspection (FAI) or CoC reports. T0 parts can also be helpful as a first iteration to test fit and function and see what exact tolerances need to be further tuned.

• Mold Flow Analysis

  • The vendor will also send you a mold flow analysis for you to look over. This will contain the suggested temperature, injection and clamping pressure, gate locations, ejector pin locations, predicted mold flow, etc. It is critical to look through this and make sure nothing functional is impacted.

  • Also of note is the expected deformation the part is likely to exhibit, as this needs to be accounted for via structural simulation and/or tolerancing.

Upload a CAD file.

Get an instant quote.

Route the work to a supplier.

Take a cut.

Repeat forever.

On paper, rapid manufacturing looked like one of those beautiful internet businesses that should have worked absurdly well. It solved a real pain point. It had software in the loop. It touched a massive industrial market. And it sat right in the middle of one of the oldest, clunkiest workflows in engineering: emailing drawings around and waiting for shops to get back to you.

And to be fair, these companies were not wrong about the problem. The problem was and is very real. Engineers do want instant quotes. Procurement teams do want a cleaner workflow. Startups do want prototypes and low-volume parts fast. The software mattered. The automation mattered. The marketplace layer mattered. The issue is that making manufacturing easier is not the same thing as making manufacturing a great business. That is where the whole sector started to wobble.

The cleanest way to understand the history of the space is through three companies: ProtoLabs, Xometry, and Fictiv. Each one represents a slightly different attempt to modernize custom manufacturing. Each one did something genuinely impressive. And each one, in its own way, ran into the same ugly truth: rapid manufacturing is operationally heavy, margin-constrained, and much harder to scale profitably than it looks in a pitch deck.

Protolabs: The Company That Proved The Category Was Real

If there is a founding success story here, it is Protolabs.

Protolabs was founded in 1999 and built a real, durable business around digitally enabled custom manufacturing. It did not start with some giant “marketplace for everything” vision. It started with a narrower and more controllable model: use software and automation to quote and manufacture prototype and low-volume parts quickly. Over time it added more processes, more geographies, and eventually a broader outsourced network. That is important. Protolabs proved this category was not fiction. Engineers really would upload files, buy custom parts online, and come back if the experience was good enough.

The good news for Protolabs was that the model worked well enough to become a real public company. The less-good news was that the limits of the model also became visible over time. Early on, Protolabs explicitly disclosed both new and existing customers. No one does this anymore as it highlights the biggest issue in the industry, CHURN. They stopped reporting this in 2017 and the rest of this analysis assumes that 40% churn continued, and that the other players experienced similar numbers. Note that this is a big assumption which may not hold, but it is without question that you can see the churn flatline from 2013-2017 between 35% and 45%, and we have no significant reason to believe that the underlying problems inherent in the industry were addressed for them or anyone else based on CAC/revenue/profit over time.

That makes sense if you think about how engineers actually use rapid manufacturing. A team prototypes a part, tests it, revises it, orders again, and then either kills the project or moves it to a different supplier for production. Great for near-term revenue. Not necessarily great for beautiful long-term customer economics. Protolabs could still make money because its model had more control and better margins than a pure marketplace, but even there the cracks showed. From 2016 to 2024 the revenue per customer stabilized right at $9,000 and GP/customer shrank from $5,300 to $4,300. Given the Customer Acquisition Cost (CAC) was $4,000, the payback period for a customer was a full year and the marketing spend was already going into the churn problem so gross profit flatlined from 2018-2024.

There is also a more practical complaint you hear from engineers all the time with Protolabs: the parts are fast, but the tolerances are not always as magical as the marketing makes them sound, and the pricing can be eye-watering. That is not really a surprise. If you build your whole brand around speed and convenience, somebody is paying for that speed. Usually the customer. And when the job gets even slightly weird, cosmetic, or tolerance-sensitive, “good enough for a quick-turn prototype” and “actually great manufacturing” can be two very different things.

Protolabs ended up being both the strongest proof that digital manufacturing is real and one of the strongest proofs that this is not SaaS. The company built a strong, useful, defensible service. It just did not turn into some magical compounding software machine.

Meet our sponsor, Jiga! They connect you directly with manufacturers. Real quotes, from real humans, for real hardware, fast.

Try Jiga Today

Xometry: Bigger Vision, Harder Economics

Protolabs proved the category was real, but Xometry tried to build it bigger.

The vision was more ambitious from the start. Instead of relying on owned manufacturing capacity, Xometry positioned itself as an AI-enabled digital marketplace connecting buyers and sellers across a fragmented manufacturing ecosystem. In its own words, buyers upload CAD files, Xometry prices them using its instant quoting engine, and the platform routes work to “vetted” suppliers while growing buyer and supplier activity on both sides. It is a very compelling pitch because it promises a business that gets stronger with scale: more buyers attract more suppliers, more suppliers improve pricing and availability, and the resulting flywheel pulls in more buyers.

And Xometry really did scale. By the end of 2024, it had 68,267 active buyers, up from 55,325 in 2023, and 4,375 active suppliers, up from 3,429. Those are not vanity numbers. That is a serious platform. It also kept widening the offering with supplier services, marketing services, financial products, and software. On the surface, this looks exactly like what you would want from a digital industrial platform.

But the filings are also unusually honest about how expensive that growth is. Xometry says outright that it will continue investing in AI, machine learning, cloud infrastructure, international expansion, supplier development, and sales and marketing, and that those investments occur before benefits are recognized, if they are recognized at all.…That sentence is doing a lot of work. It is a polite SEC-filing way of saying the platform still needs a lot of “fuel.”

Xometry hasn’t seen the Protolabs plateau yet, but it’s on the way. When your revenue per customer is $7-9k/year and CAC is $5-6k with 28% gross margins there is a payback period per customer of ~2.5 Years!

There is also a product problem buried inside the model itself. One of the biggest complaints engineers have with Xometry is the black-box nature of the quoting and sourcing flow. You upload a part, get a number, and in many cases have no idea who is actually making it. On two different orders, the answer might be two completely different suppliers with two totally different capabilities, surface finish standards, communication styles, and tolerance discipline. That is great for marketplace flexibility. It is less great if you are trying to build confidence that the fifth order will look and feel like the fourth one. The platform abstracts away the supplier, which is part of the convenience, but it also abstracts away accountability in a way that can make repeatability feel shakier than customers want.

Fictiv: Strategic Logic, Ugly Short-Term Math

Fictiv is probably the clearest example of how ugly the model can get. It raised about $188M, was reportedly valued around $1B+ at the peak, and sold to MISUMI in 2025 for $330M. Once MISUMI started consolidating Fictiv in July 2025, the FA segment (where Fictiv was added to within Misumi) immediately showed the pattern you would expect from a low-margin acquired marketplace business: sales up, profit down.

Misumi released their Financial 2025 year in review document here where they have specific breakouts for Fictiv business performance as shown in the graphic below. $98M in revenue, -$20M in operating income… Would you pay $330M for that business?

Though their operating margins improved as a result of the consolidation, -21% operating margins is still shockingly terrible. If you factor in the $27M spent on “M&A associated costs” (probably lawyer fees) and goodwill amortization (usually a sign that the reported acquisition optics are rough), Fictiv lost Misumi almost $50M in Y1 post acquisition.

To quote Misumi:
”As with sales, Pre-Fictiv consolidated profitability increased due to improvements from the initial plan. Post-consolidated profits declined due to impact of Fictiv acquisition.”

The miracle here is that they managed to be acquired at all with -34% operating margins the year prior. Based on these numbers, my presumption is that Fictiv follows a very similar trend to Xometry and Protolabs: high CAC (customer acquisition cost), high churn, so their best way to juice revenue was to increase marketing spend to acquire customers faster than they churn.

I am curious to see if the synergy plan Misumi has developed of accelerating Fictiv growth by leveraging Misumi factories and manufacturing partners while cross selling to existing Misumi customers pans out. The trendline is good given the substantial improvements in operating margin already since joining Misumi, and Misumi is “aiming to achieve Fictiv standalone profitability, no later than FY27.”

Why Jiga Might Actually Fix Rapid Manufacturing

The history of this space makes one thing pretty clear: the problem was never demand. Engineers absolutely want faster quoting, clearer communication, and a better way to buy custom parts. The problem was that too many companies tried to force manufacturing into a pure software pattern. They optimized for speed at the top of the funnel, but left too much ambiguity, too much supplier opacity, and too much operational mess underneath. Jiga’s approach looks different. Jiga explicitly does not do instant quotes, instead offering human-reviewed quotes in 24 hours and direct communication with vetted manufacturers. In other words, it is built around the reality that good manufacturing is still a relationship business, and the software should support that rather than pretend to replace it.

That matters because the ugliest economics in this category usually come from the same handful of problems: quoting mistakes, bad handoffs, poor supplier fit, expensive exceptions, and customers who churn once the job gets more complex. Jiga seems designed to attack those exact weak points. Its pitch is not “throw a CAD file into a black box and hope for the best.” It is closer to: keep the manufacturer in the loop, keep the communication visible, use AI to remove the administrative sludge, and make sourcing feel like an extension of the engineering team instead of a totally separate function. The company describes itself as handling quoting and communication in one place, using AI workflows for the repetitive admin work, and giving visibility from prototype through production. That is a much more believable way to improve margins and retention than just chasing more RFQs with more automation.

On the downside, Jiga’s supplier network is still considerably smaller than the established players, a natural constraint for a younger company that vets manufacturers more selectively. On the bright side, the suppliers listed do seem to have a broad repertoire of possible manufacturing processes and work they are capable of, and are very geographically diverse (though strong emphasis is on USA and Chinese vendors). ITAR/ISO work is possible when needed.

Story time: When I was co-oping at Relativity back in 2020, I was one of 4 interns. I was on the propulsion team which exclusively hired PhDs at that point, so they pretty much thought I would be useless for anything more than grunt work until I proved them wrong. As such, the first task they gave me was to source ~30 different parts for the rocket gimbal subassembly, some COTS, some custom (DMLS, heat treat, wire EDM, CNC). Not as illustrious as I thought it would be, but boy did I tackle it with gusto (once convinced a bolt supplier to sell me bolts below cost by pitting them against a rival supplier), and I can say with absolute certainty it would have been way simpler if Jiga existed at the time. Instead, what I did was reach out to about 50 suppliers with emails and calls and dozens of back and forths in order to obtain a few quotes. Something you don’t learn until it happens, is that when a supplier later makes a mistake, it’s the good ones that tell you, but either way you basically have to work with it unless you want to purchase more raw material, have an abundance of timeline (literally never happened), and kickoff the project with a whole new vendor from scratch. The easiest way to tell quality was actually to visit each vendor in person to see how the sausage is made. Luckily I just wanted to be helpful and didn’t mind visiting every shop in a 50 mile radius for the learnings alone, but I definitely wish Jiga had been around at that time for me to use; having easy access to vetted suppliers would’ve let me handle way more of the fun stuff sooner by preventing me from getting bogged down in the sourcing swamp.

Unlike Protolabs and Xometry, Jiga is not a public company, so they do not disclose full financials, but the company does report an annual activated customer churn rate under 11%, an extremely stark contrast to the ~40% churn rates seen historically across the category. Additionally, in late 2025 they announced a $12M Series A led by Aleph, with participation from Symbol and Y Combinator, and on its own materials the company says it is cashflow positive and growing fast. That does not guarantee success, but it is a very different starting point from the “raise huge, subsidize growth, figure out economics later” playbook that has burned so many others in this category. If Jiga can keep doing what it appears to be doing — using software to cut friction, using people where judgment still matters, and helping teams scale from prototype into production without losing clarity — it has a real shot at solving the exact problems that made the first generation of rapid manufacturing platforms so hard to love.

Try Jiga!

Note: Though this post was sponsored by Jiga, we do honestly believe that the rapid manufacturing industry needs to change and that Jiga has a chance to improve real problems with sourcing.

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

Meet our sponsor, Jiga! They connect you directly with manufacturers. Real quotes, from real humans, for real hardware, fast.

Try Jiga Today!

🦸 Guest Writer: Madhav!

This week’s newsletter is a special edition written by Madhav Bhat on softgoods manufacturing! Madhav graduated from the University of Michigan with a BSE and MSE May of 2024. Throughout his recruiting journey he interviewed for mechanical engineering internships and full-time positions with companies from large Fortune 500 corporations (think Ford, Michelin, Apple) to growth-stage companies (think Tesla, SpaceX, Anduril) to startups ranging from 3-1000 people in size working on all sorts of projects.

Introduction

When I used to work in automotive manufacturing, things like tolerance stack-up analyses and GD&T were pretty fleshed out for the rigid bodies involved. Material properties were known and you could get a pretty good idea for the process windows needed for the parts you were working with. Recently, I got some clothes I ordered that were different from the online size chart and it made me think, what does process control and tolerancing look like for softgoods? Like clothing? Not only from a manufacturing standpoint but down to design - is there CAD software for apparel? Actually, forget that. End-to-end how is clothing made, especially at the scale we’re at today? This made me dive into a rabbit hole to answer “how does mechanical engineering work for apparel and other softgoods?”.

Design and the “tech pack”

Let’s start with design. The universal term for a design summary for softgoods/apparel is a “technical design pack” or “tech pack” for short. The tech pack consists of:

• Technical Drawings & Callouts

  • Precise line drawings with notes on construction details, stitching, seams, hardware, and placement.

• Bill of Materials (BOM)

  • A complete breakdown of all materials and components. It includes fabrics, foams, fasteners, and sensors required to build the product.

• Measurements & Grading

  • Dimensions, tolerances, and size variations to ensure consistent fit across different body types or product sizes.

• Assembly Instructions

  • Step-by-step construction methods that guide how pieces come together, whether sewn, bonded, or mechanically fastened.

• Testing & Performance Standards

  • Requirements for durability, washability, strength, or medical-grade compliance, depending on the product category.

• Labeling & Branding

  • Placement of logos, care instructions, or certifications that connect the product to its brand identity and compliance needs.

Honestly this seems pretty similar to traditional product development however let’s dive deeper into parts 1 and 3, the technical drawings and measurements.

Technical Drawings & Measurements

In traditional mechanical design, this is pretty easy to visualize. You can use any CAD tool to create engineering drawings of the part/assembly you want to create with the appropriate tolerances and measurements listed. Simple GD&T.

It’s even simpler in apparel design and it looks very similar to a sizing chart you’d see online. The common tool for creating these drawings is Adobe Illustrator.

Looking at this, my immediate reaction was “no wonder there’s so much variance this seems so vague”. Also where are the measurements?” After some more digging, there’s actually 3 documents that come together to paint the picture of how apparel should be designed. There’s the flat which is what is pictured above, there is the POM diagram which contains the measurement locations (what is actually being measured - pictured below), and there’s a spec sheet table which contains the actual numbers (also pictured below).

Ok so this makes sense, but where are these tolerances coming from? The primary driving factor is something called “the grade rule” which is pretty simple: the difference in each dimension between each size divided by 2 is generally what you should set your tolerance to. Example:

You have a t-shirt in size M with a chest measurement of 20”. Size L is 21”, and size XL is 22”. That 1” jump between each size is the grade rule for chest width.

Now, what tolerance do you set on the chest measurement? You can’t allow more than ±½” — because if you did, a sloppy M could measure 20.6” and a sloppy L could measure 20.4”, and suddenly your M is bigger than your L. The sizes have crossed over and your sizing system has collapsed.

So the tolerance is constrained by the grade rule. If sizes jump by 1”, your tolerance ceiling is ±½” (half the grade). If sizes jump by 2”, you could allow up to ±1”. Makes sense. There are also of course other factors that contribute like the process capability of the factory (similar to traditional mfg.), size of the details (ex. width of a skirt vs. side pocket of it), material properties, and brand standards. But the grade rule is the general loose rule.

The final product here is a design for a garment that can be tested to make sure it can be productionized.

Simulation and Testing

I was surprised to learn that there were simulation tools built around the garment industry and they actually measure similar metrics to what tools like FEA do. One specific example I’ll focus on is CLO 3D. The simulation software requires a lot of input on the material properties of each of the elements. Once you plug in the material properties, there’s an interesting way to do the validation and make sure it’s correct:

CLO validates its digital fabric twins using a drape test: physical and digital fabric are cut to the same 50×50cm sample and draped over an identical cylinder. The digital twin is considered accurate if the height of each drape corner matches the physical sample within 1cm, and the number of drape flares matches within one.

Once calibrated, the designer imports 2D pattern pieces, virtually sews them together, and drapes the resulting garment on a parametric avatar. The software then runs a physics simulation and outputs several analysis modes:

• Stress map — shows force per unit area along the fabric grain, displayed as a color gradient from blue (no stress) to red (high stress). Tight areas around armholes, crotch, or bust show up immediately.

• Strain map — shows how much the fabric is stretching relative to its natural state, which tells you whether a stretch fabric is being deformed beyond its design range

• Pressure map — shows external pressure the garment is exerting on the avatar body, critical for compression garments and sportswear

• Fit map — a simplified pass/fail overlay showing where the garment has sufficient ease, is too tight, or is too loose

Once simulation is done, we can move on to physical testing, a lot of which needs to be in accordance with regulations set by governing bodies like ASTM International, ISO, and AATCC (American Association of Textile Chemists and Colorists). There are some pretty interesting tests here including:

• Seam strength is evaluated by sewing a seam in the fabric, clamping it at one end, and pulling with weights at the other until the seam separates a specified distance. Similar to a standard tensile test using an Instron.

• Dimensional stability tests use ISO 5077 and ASTM D4970 to assess shrinkage or elongation after laundering. Acceptable shrinkage rates generally range from 3–5% for most adult apparel and as low as 1–2% for fitted items.

• Colorfastness to washing (AATCC Test Method 61) simulates multiple home launderings to assess color change and staining. Colorfastness to light (AATCC Test Method 16) evaluates resistance to fading from light exposure. Colorfastness to rubbing (ISO 105-X12) tests whether color transfers to other surfaces under friction.

• Chemical / Safety Tests

  • Formaldehyde content — residual from wrinkle-resistant finishes, regulated in children’s wear

  • pH testing — ensures fabric won’t irritate skin

  • Heavy metals — particularly lead in dyes and prints

  • Restricted substances — phthalates, azo dyes, flame retardants, etc.

  • Flammability — mandatory for children’s sleepwear and certain workwear

  • Much more test based on the type of apparel being made (waterproofing, moisture wicking, thermal resistance, etc.)

Overall the guiding principles around testing and simulation are pretty analogous to what we learn in class for testing in mechanical engineering. Now that we’ve tested and meet all the regulations to bring these clothes to the market, it’s time to make them!

Subscribe to Pipeline Media Lab and get the free PDF: 5 Engineering Decisions that Built Careers. Hamilton. Moore. Grove. Real decisions, real stakes, frameworks you can use this week. Scan the QR code or subscribe here.

Manufacturing

So now that we have a design spec for a piece of clothing, what happens next? Assuming it’s going to be manufactured at scale, it’s going to go through a couple steps for industrial garment production.

The general flow is: design → fabric sourcing → pattern making → grading → marker making → fabric spreading → cutting → bundling → sewing → in-line inspection → ironing and finishing → tagging → final inspection → packing → dispatch. Almost all of this is manual.

To be honest, the majority of clothing factories have zero automation. The biggest automation here is the fact that they’re not hand sewing or cutting with scissors. Even a company like Shein that is producing clothing at en masse (1.2 million items of clothing per day) is not really using any manufacturing automation, just a bunch of small factories (and some complicated software).

For the minority of apparel brands that are investing in manufacturing automation, some of the key areas are automatic cutting machines, and minor automations in pressing, tagging, folding and cutting. There are also machines that do automatic sewing, but not all sewing. The big problem is general-purpose automated sewing of limp fabric. Every automated sewing machine on the market today either requires the fabric to be clamped into a rigid fixture, or requires a human hand to pick up each panel, align the edges, and feed the material through at the correct tension.

The gap between current machines and the ideal comes down to three things humans do by feel that robots can’t yet do reliably at production speed: picking a single fabric ply off a stack without distorting it (the material has no rigid reference geometry), tracking a seam edge under tension as the fabric stretches and drifts (requiring constant micro-corrections a skilled sewer makes unconsciously), and navigating 3D curved seams like setting a sleeve into an armhole, which requires easing two deformable surfaces into alignment simultaneously.

Until those three problems are solved cheaply enough to beat a $1,200/month sewer in Guangzhou, the sewing floor will look essentially the same as it did in 1965 (but with some of the machines I mentioned above). Maybe humanoids are the answer here (but even they are not dextrous enough yet it seems like to take the job of a skilled sewer), but if you’re looking to be a founder, this is a pretty interesting problem with a large TAM!

One thing to emphasize here is that a lot of apparel companies don’t own this manufacturing workflow end-to-end. They’ll primarily be responsible for just the design and will outsource the manufacturing to some contract manufacturer. This is pretty different from what we see in traditional mechanical engineering.

However, the ones that are vertically integrated and do it well can reap some big benefits. A big case-study is Zara that owns their factory and can produce garments in response to sales data and market trends pretty easily. They also apparently use a technique called “modular manufacturing” where garments are broken down into components that can be assembled later (so likely fewer total steps needed to assemble a completed garment).

Also wow, the infrastructure they’ve built around this is insane. Zara has a 5 million sq. foot distribution center in Spain called “the Cube” that is directly connected to 11 Zara-owned factories via 200 km of underground tunnels. With monorail. Like what. Maybe we should do an entire article on just Zara.

Anyways, that’s a high-level overview of how apparel is made! To be honest this is pretty dated and focuses a lot on how apparel has been made in the past. It also probably heavily depends from company-to-company, and I’m sure giants like Nike and Adidas have much more processes and advanced technology here that I didn’t capture in the article. There are also some startups that are really revolutionizing how clothing is made (the main one I’m tracking right now is Unspun) and I want to capture how they’re changing the processes described above. Maybe for a future blog!

My desired goal state for apparel manufacturing is the following:

This week’s newsletter is a special edition written by written by Brad Hirayama of Pipeline Media Lab in collaboration with Jim Berry. Brad is a biomedical engineer and avid technical communicator. If you like this content, check out some more of his content here!

⚙️ Mechanical Engineering Resources:

We have put a dozen guides for mechanical engineering students and early professionals on our website.

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it take to be The 10X Intern

👶 Meme Of The Week

The complete guide to DFM and Assembly

Jiga, a platform that connects hardware teams with vetted manufacturers, no black box between you and the people making your parts, put together a guide on the 7 DFM principles that should underpin every design decision. Covers why manufacturability needs to happen early, not as a final check, and how to work with suppliers before your leverage disappears.

Check Jiga Out!

You’ve Been Learning GD&T Backwards

Most engineers start with the symbols. The ones who actually use it start with function.

Written in partnership with Jim Berry, Technical Expert and GD&T Instructor

Every mechanical engineer remembers the first time they encountered a fully called-out GD&T drawing. The little boxes stacked with geometric characteristic symbols, tolerance values followed by datum letters, modifiers they’d never seen, and a datum reference frame that seemed to materialize from nowhere. Most engineers do exactly what anyone would do: find a symbol chart, memorize what each icon looks like, and hope that institutional knowledge carries them the rest of the way.

Years later, many are still faking it. They copy tolerance callouts from previous drawings without knowing if the values are appropriate. They assign datums alphabetically rather than functionally. They write drawings that pass design reviews and then generate expensive phone calls from the shop floor.

The problem is not the symbols. The symbols are just notation. The problem is that GD&T is almost universally taught as a symbol language when it is actually a functional specification language - one that starts not with a drawing but with a question: how does this part behave in the assembly? If you skip that question, every callout you write is a guess dressed up as engineering.

This article walks through the framework that makes GD&T systematic rather than intuitive. It will not replace the ASME Y14.5 standard. What it will do is give you a mental model that makes every drawing you read - and every drawing you produce - immediately more deliberate.

Why GD&T Breaks Everyone’s Brain

A professor introduces GD&T by walking through the symbol families: form controls (flatness, straightness, circularity, cylindricity), orientation controls (parallelism, perpendicularity, angularity), location controls (true position, concentricity, symmetry), and profile controls. Students learn what each symbol looks like and what geometric characteristic it governs. Then they move on to the next chapter.

Nobody shows them a real assembly and asks: given how this part functions, which of these controls belong on this drawing, and why? Nobody explains that a flatness callout on a mounting surface exists to prevent the part from rocking against its mating face, or that a position tolerance on a hole pattern exists to guarantee the fasteners can always be installed regardless of where individual holes land within their tolerance zone. The callouts are presented as a catalogue of options rather than answers to specific functional questions.

The result is engineers who can identify a perpendicularity callout but cannot tell you whether it is necessary, whether the tolerance value is correct, or what happens at the shop floor if it is wrong.

There is also a second, more practical source of confusion: the feature control frame itself. That rectangular box that is divided into compartments and appearing next to almost every GD&T callout (shown in figure 1), looks arbitrary until you understand its structure. The frame is read from left to right and contains:

• Geometric symbol -> What is being controlled

• Tolerance zone shape -> What is the shape of the zone being controlled

• Tolerance zone size -> What is the size of the zone

• Material Modifier -> Whether your tolerance is fixed or variable

• Datum References -> Inspection set up in order of precedence

Engineers who treat the frame as a black box to produce drawings that are technically valid but functionally wrong.

Figure 1. Read from left to right, the anatomy of the feature control frame can be broken down into these 5 pieces.

The third source of confusion is coordinate tolerancing. Most engineers learned to dimension parts with plus-or-minus linear tolerances on X and Y locations. That produces a square tolerance zone. GD&T position tolerance produces a cylindrical tolerance zone. For the same stated tolerance value, the cylindrical zone is approximately 57% larger in area than the square zone. That is the difference between a part passing and failing inspection, and it has nothing to do with whether the part actually fits. Engineers who do not understand this distinction routinely reject good parts or accept bad ones.

The Only 4 Things GD&T Controls

Before the framework, one foundational fact that most engineers are never told explicitly: GD&T controls exactly four geometric characteristics (shown in figure 2). Everything else is notation. Every symbol in the standard is controlling form, size, location, or orientation. That is the complete list. When a drawing looks overwhelming, running each callout through this filter immediately clarifies what the engineer was trying to control and whether the right tool was used to control it.

Figure 2. The only four characteristics GD&T controls.

Form

Form describes how straight, flat, round, or cylindrical a surface actually is, independent of any reference to other features. A wall that should be perfectly flat has some bow to it. A hole that should be a perfect cylinder will have some variation in its cross-section as you move along its axis; which means it might be slightly egg-shaped, tri-lobed, or barrel-shaped within its allowable size range. Every manufactured surface has form error. GD&T lets you define how much is acceptable before it affects function.

The four form controls (shown in figure 3) are flatness (surfaces), straightness (line elements or axis), circularity (cross-sections of cylinders or cones), and cylindricity (the full cylinder surface simultaneously). Critically, form controls never reference datums — form is an intrinsic property of a single surface, not a relationship between features. If you see a form control with a datum reference, something is wrong with the drawing.

Figure 3. Form variation in a cylindrical feature - Nominal, Oval, Tri-lobed, Barrel

Size

Size defines how large or small a feature can be within its acceptable range. For any feature where two surfaces or elements oppose each other - a hole and a pin that must mate, a tab that must fit in a slot, a block that must slide into a channel - you must specify the allowable size range. The largest pin must always fit in the smallest hole. The smallest pin in the largest hole must not have so much clearance that it rattles or loses alignment.

Size is controlled with coordinate tolerancing: a nominal dimension followed by a plus-or-minus value. A pin called out as 19±0.5 can range from 18.5mm to 19.5mm. A mating hole called out as 20±0.5 can range from 19.5mm to 20.5mm. At their worst-case combination (largest pin (19.5) into smallest hole (19.5)) the pin barely fits. That calculation is not optional. It is the minimum verification required before the drawing leaves your desk.

Location

Location defines where a feature is relative to its functional reference, usually its datum reference frame. A hole can be perfectly round and exactly the right size and still be in the wrong place. Location controls answer the question: how far off from its nominal position can this feature be and still allow the assembly to function?

This is where coordinate tolerancing and GD&T position tolerance diverge most significantly. A plus-or-minus X/Y location tolerance creates a square tolerance zone: the hole axis must fall within a box defined by the bilateral tolerances in each direction. A GD&T position callout with a diameter symbol creates a cylindrical tolerance zone centered on the true position of the hole. For a stated tolerance of ±0.5mm in coordinate tolerancing, the square zone has an area of 1.0×1.0mm = 1.0mm². The equivalent cylindrical zone with a diameter of 1.0mm has an area of π(0.5)² ≈ 0.785mm² - roughly 57% more usable tolerance area - while still guaranteeing the same assembly performance as the square zone. More tolerance, no functional compromise.

Orientation

Orientation controls how tilted a feature is relative to a datum reference. A hole axis that should be perpendicular to a mounting face but has angular deviation. A surface that should be parallel to another but has a slight tilt. Orientation always references at least one datum because tilt is only meaningful relative to something else.

The three orientation controls are perpendicularity (90° nominal), parallelism (0° nominal), and angularity (any other angle). Each defines a tolerance zone - two parallel planes, or a cylinder - within which the surface or axis must fall when the part is oriented relative to the specified datum. Every datum reference in a feature control frame is, at minimum, an orientation control. Location controls are always orientation controls as well, because you cannot specify where something is without implicitly constraining how it is tilted.

Five engineers made one decision each that defined their careers. Margaret Hamilton, Gordon Moore, Andy Grove. The frameworks they used are in a free PDF - yours when you subscribe to Pipeline Media Lab at this link or by scanning the QR code.

The Function-First Framework

With the four characteristics as a foundation, here is the framework that makes GD&T systematic. Three steps, applied in sequence every time, to every part.

1. Understand the assembly before you touch a drawing

The first thing you do when approaching a part is not open the drawing. It is to understand how the part interacts with its mating features. Where does it sit? What surface does it press against first? What locates it in-plane? What prevents it from rotating? What has to fit inside it or around it? Take the square block example shown below.

Figure 4. The square block example from Jim’s Webinar.

This is a square machined block - a fixed thickness, a centered hole at ⌀20mm, and a chamfer on one corner that acts as a poka-yoke, forcing the part to install only one way. Before a single tolerance is written, the only question that matters: how does this part sit in the assembly?

The block seats down on a flat mounting surface first - that’s datum A, the primary reference. It then slides laterally until it contacts a side surface - datum B, the secondary. Finally it slides in the perpendicular direction until it contacts a third surface - datum C. Three surfaces. Three points of contact. Three datums.

Notice what this does to the feature control frames. Datum A gets a flatness qualifier; which means if the bottom surface rocks, the entire reference frame is unreliable. Datum B gets a perpendicularity qualifier back to A - the side surface must be controlled relative to the primary before it can be used to locate anything else. Datum C gets the same treatment. Each datum qualifies the next one. The callouts build on each other in the same sequence the part physically seats.

The hole is the last thing specified. Its size tolerance (⌀20 ± X, left as a placeholder until a tolerance stack is calculated) controls form and size. Its position callout references A, B, and C in that order, controlling orientation - the axis must be perpendicular to A and parallel to B and C - and location, 50mm out from datum planes B and C. The chamfer goes to the default profile tolerance; no specific callout needed because the poka-yoke geometry prevents mis-installation regardless.

2. Classify every feature as a surface or a regular feature of size

Once you understand the assembly, look at every feature on the part and ask a single question: can I grab this with a caliper and get a repeatable measurement between two opposing points or surfaces? If yes, it is a regular feature of size. If no, it is a surface.

Cylindrical features of size are the most common: holes and pins. The inner diameter of a hole and the outer diameter of a pin are both features of size: you can grab either with calipers and get a direct size reading. Planar features of size are equally important but less frequently recognized: a block that must fit into a slot, or a tab that must fit into a notch. If you can grip both sides of the opening or both sides of the tab simultaneously, it qualifies as a regular feature of size (shown in figure 5).

Figure 5. The caliper test, an easy way to identify regular features of size.

Most engineers blur this distinction and end up with drawings where surfaces get size tolerances applied incorrectly, or features of size are controlled with profile callouts that don’t communicate what was intended. The result is a drawing your manufacturer has to interpret rather than read.

3. Assign datums in functional order - A, B, C maps directly to how the part seats

Datums are not arbitrary labels. Datum A is the most important mating interface for your part in the assembly. Datum B is the second. Datum C is the third. The sequence describes the physical order in which the part contacts its mating features.

For a bracket that mounts to a wall: datum A is the flat back surface (the part seats against the wall here first), datum B is established by the pattern of two fastener holes (these locate the bracket in-plane once it’s against the wall), and datum C is the third constraint. The inspection setup follows exactly: set the part against A, then locate against B, then check everything else relative to that reference frame.

A flat primary datum surface (datum A) constrains three degrees of freedom: translation perpendicular to the surface, and rotation about two axes parallel to it. A secondary datum feature - often a hole pattern or a cylindrical surface - constrains two more: translation in the two remaining in-plane directions. A tertiary datum constrains the final degree of freedom: rotation about the axis perpendicular to the primary surface. Six degrees of freedom total. Three datums, in functional order, lock the part completely.

The datum reference in any feature control frame tells the inspector exactly how to set up the part for measurement. The sequence is the assembly process, translated into inspection language. Once you see it this way, datum A–B–C stops being alphabet soup and starts being a precise description of how your part lives in the world.

One nuance that trips up even experienced engineers: many datum features need a form qualifier before they can serve as a reliable reference. A flat surface called out as datum A should also carry a flatness callout - qualifying that surface to be flat within some tolerance before it establishes the primary datum plane. A surface that rocks by 0.3mm cannot reliably establish a plane. The qualifier ensures the datum feature is good enough to be a datum.

The Interview Question

Companies that are bringing on mechanical engineers, especially in consumer electronics, robtics, aerospace, to name a few, test GD&T in almost every technical interview. The question will usually look something like this (shown in figure 6) and you’ll be asked:

Here’s a wall bracket. It mounts to a surface, holds a divider panel in a central opening, and is secured by two fasteners through clearance holes. Walk me through how you would fully define it with GD&T.

Figure 6. A Mock interview question based on a bracket mounted to a wall.

The easy answer is to just call out symbols based on what you see. “I’d put a flatness callout on the mounting surface, true position on the holes, and a size tolerance on the opening.”

Symbols listed. No rationale. No tolerance values. No explanation of why the datums are in the order they’re in, or how you’d know if any of those numbers were right. Any engineer who has skimmed a GD&T textbook can produce this answer. It tells the interviewer nothing about whether you could actually spec a real part.

Here’s how you should answer this problem:

Before I write anything, I want to understand the assembly sequence. This bracket seats its flat back surface against the wall - that’s the first and most critical interface, so that surface is datum A. I’d add a flatness qualifier of 0.1mm to it. If the surface is more than 0.1mm out of flat, the bracket will tilt or rock against the wall, and the divider panel inside the opening won’t sit perpendicular. The flatness qualifier is protecting the function of everything else downstream.

The two fastener holes are next. Once the bracket is seated against the wall and the fasteners are torqued, those holes locate the bracket in-plane: they constrain it from translating or rotating in the plane of the wall. Both holes together establish datum B as a pattern, not individually. I’d add a position callout to control their perpendicularity to datum A and the spacing between them, with the tolerance value left as X until I’ve done a proper tolerance stack. You don’t put numbers in until you know what the assembly can actually absorb.

The divider panel opening is a regular feature of size - I can grip both walls of it with calipers, so it gets a size tolerance, not a profile callout. If the panel is 24.5 ±0.2mm wide, the opening needs to be 25 ±0.2mm. Worst case, the largest panel is 24.7mm and the smallest opening is 24.8mm: a 0.1mm clearance, the panel installs. That calculation has to be done before the drawing ships, not after first articles come back wrong.

The center plane of that opening also gets a position callout tied back to A and B. This controls where the opening actually sits relative to the fastener holes — it ensures the panel lands where it’s supposed to in the assembly, not just that it’s the right width. Finally, the two smaller holes that secure the panel get their own size tolerances and position callouts, also referenced to A and B. Everything traces back to the same datum reference frame.

Figure 7. Fully dimensioned with datums and GD&T tolerancing.

That answer demonstrates five things simultaneously: assembly awareness (datums assigned in functional order, not alphabetically), feature classification (the opening identified as a regular feature of size before any callout is chosen), worst-case analysis executed on the spot with real numbers, a complete datum reference frame that mirrors the inspection setup exactly, and the discipline to leave tolerance values open until the math supports them. Any one of those signals competence. All five together signals someone who has actually built hardware.

The Bonus Question: Unlocking MMC

You applied a position callout to the fastener hole pattern. What happens to that tolerance zone when you added an M modifier?

To answer this question, you have to go back to the bracket. The two clearance holes establish datum B: they locate the bracket against the wall once the fasteners are torqued. The position callout controls their perpendicularity to datum A and the spacing between them. Now you add the M modifier to that callout and you’re calling out the maximum material condition.

Maximum Material Condition means the stated tolerance applies when the hole is at its smallest allowable size or the most material remaining, the tightest condition. For a clearance hole, that is its narrowest diameter. As the hole grows larger - departing MMC toward Least Material Condition - it earns bonus tolerance equal to that departure. The position zone expands.

Here’s the math: Say the clearance holes are called out at ⌀14 ±0.2mm. MMC is ⌀13.8mm. The stated position tolerance at MMC is ⌀0.5mm. A hole produced at ⌀14.0mm has departed MMC by 0.2mm - it gains 0.2mm of bonus tolerance, bringing its total position zone to ⌀0.7mm. A hole produced at LMC of ⌀14.2mm gains 0.4mm of bonus, for a total zone of ⌀0.9mm. The fastener still fits - a larger clearance hole gives the bolt more room to find its way in. More hole, more forgiveness, same functional result.

This is why the M modifier exists. Without it, every hole is held to the same ⌀0.5mm position zone regardless of produced size. With it, larger holes - which are easier to assemble and naturally more forgiving - get a larger zone. The parts that should pass, pass. The ones that shouldn’t, don’t. Manufacturing yield goes up without touching the functional requirement.

Now connect it back to the divider panel opening. The slot was called out at 25 ±0.2mm against a panel of 24.5 ±0.2mm. The position callout on the slot’s center plane was referenced back to A and B. You could apply an M modifier there too: a wider slot opening has more room to accommodate the panel even if its center plane is slightly off. The functional requirement is that the panel goes in. A 25.4mm opening that’s 0.3mm off-center still accepts a 24.7mm panel cleanly. A fixed position zone wouldn’t know that. MMC does.

Figure 8. MMC Bonus Tolerance: Position zone grows as hole grows.

GD&T is not a documentation formality. It is a precision communication protocol between the engineer who specifies a part and every person who has to make it, inspect it, and assemble it. A drawing that uses GD&T correctly eliminates ambiguity at every step in that chain. A drawing that uses it incorrectly or uses it as decoration on top of coordinate tolerancing, creates ambiguity that compounds downstream, from first articles to production ramp.

The engineers who build hardware that works the first time are the ones who asked the right question first: how does this part behave in the world? Start there. The symbols will follow.

This week’s newsletter is a special edition written by Madhav Bhat based on an interview with Ultrahuman! Madhav graduated from the University of Michigan with his BSE and MSE May of 2024. Throughout his recruiting journey he interviewed for mechanical engineering internships and full-time positions with companies from large Fortune 500 corporations (think Ford, Michelin, Apple) to growth-stage companies (think Tesla, SpaceX, Anduril) to startups ranging from 3-1000 people in size working on all sorts of projects. If you have contacts in the wearables or humanoid spaces he can talk with please reach out to hardwareishard@gmail.com!

⚙️ Mechanical Engineering Resources:

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

What goes into making a fitness wearable?

Recently (and especially in San Francisco), I’ve been noticing more bulky rings on people’s fingers and faceless watch straps on people’s hands. People have become more interested in rigorously tracking health metrics and there are a couple companies capitalizing on that with devices in the fitness wearables market. These devices track health metrics like sleep, HRV, and steps as well as custom metrics like “body battery” or a “recovery score”.

As someone who’s tried these products before, it begs the question, how are we packing so much power into such a small form factor that has to go through some of the most rigorous everyday usage a device can go through? We decided to explore that in these week’s newsletter through some research, as well as talk with some experts like Ultrahuman Ring’s head of mechanical engineering - Mainak Mitra.

What actually makes up these devices?

Exterior

For this section, we’ll focus specifically on fitness-tracking rings (which I find super interesting) and start from the outside in. For the portion of the ring that faces outwards from the finger, these can be made from other materials, but primarily are made out of titanium since the important element here is a strong stiffness to weight ratio. You’ll see this theme of weight-saving pop-up many times (normal ring jewelry is ~3g, Ultrahuman has gotten their rings down to 2.5-4g). In addition to the titanium on the outside, there are various coatings that are used on the ring’s exterior for scratch resistance. This can include:

• Tungsten carbide carbon (WCC) coating - 2x as strong as steel but also 2x as dense

• DLC (Diamond-like-carbon) coating - very hard but also very sleek (pretty cool)

I could not find a great source for how the titanium exterior is constructed but can assume it’s either metal injection-molded or CNC machined.

On the outside of the ring that faces into the finger, there is either:

• a medical-grade epoxy resin that covers the otherwise exposed circuitry. This is what was on previous Oura ring generations and the Ultrahuman ring

• also titanium. In later Oura ring generations there is a titanium inner cover with holes for the sensors.

Interior

On the inside we have the sensor suite and core hardware which can reliably consist of a couple key components:

• Optical Heart Rate / PPG (Photoplethysmography) Sensor - this is primarily for heart rate and oxygen monitoring

• Temperature Sensor - this is to track baseline skin temperature

• Inertial Measurement Unit (IMU) — 6-Axis Motion Sensor - this is how the ring monitors if you are moving or standing still

• BLE radio - used for communication from and to device

• Inductive charging coil - used to charge the device. This is extremely unique in that it is a custom curved battery that serves as a major constraint and space-consumer in the ring.

These components are placed on a flexible PCB made of polyamide that’s curved into a J-shape before being inserted into the ring. You don’t see deliberately curved PCBs too often!

Once the PCB is made and sensors are in place, adhesive is used to mount it to the titanium shell and resin is poured over the assembly to finish out the creation of the ring (and also support in holding all the components in place). One interesting fact is that the resin requires a 24-hour cure time which is a production bottleneck.

Testing

One thing I wanted to dive deep on is how do you test something that’s constantly in contact with other surfaces all day and cannot deform?

Like other consumer electronics, the rings go through hardness and abrasion testing, as well as significant drop testing.

They also go through accelerated life-cycle tests with thermal loading. One thing Mainak pointed out that was an interesting edge-case was that “biohackers” often take their rings from the Sauna, directly into a cold plunge, so going from 150 deg F to 30 deg F in a matter of seconds, and they might be doing this every day, so the ring needs to go through rigorous thermal cycling tests to capture this!

Because the ring sits on your finger, it needs to go through a Cytotoxicity test as well to evaluate the general toxicity of the device/material in compliance with ISO 10993.

And finally one of the biggest factor of designing a fitness wearable is answering “is this comfortable??”. Ultrahuman tests rings amongst a wide range of demographics with different skin color and thickness to ensure comfort and sensor performance. These users need to cover 95% of the population.

Interesting properties

In a ring form factor, the mechanical engineer is in charge. The ring needs to be a certain dimension out of concern for comfort for the user, and this in turn defines a lot of the electrical hardware choices (ex. how large of a battery can we use?) for the ring.

Also in this small form factor, tolerance stack-up plays a huge role and there can’t be compromises. This is why Ultrahuman (and Oura I believe) manufacture most components in-house - take a deeper look in this article by Android Central! Components like flexible PCBs and curved batteries make the manufacturing process pretty bespoke and unique.

In fact, if you look into the article above, the entire manufacturing process is pretty hands-on and arduous(although this article is about a year old). Components are soldered onto PCBs by hand, resin takes 24 hours to cure and even after that needs to be polished and refined to make sure it’s comfortable on skin. Scaling is one of the most difficult aspects of designing a new product that requires meticulous detail on a very small scale.

Conclusion

What makes smart rings fascinating from an engineering perspective is that they sit at an intersection that doesn’t really exist anywhere else in consumer electronics. You need jewelry-level craftsmanship — titanium forming, surface coatings, comfort testing across thousands of finger shapes — combined with the sensor integration and firmware work of a proper electronic device.

However, once again, we are still in the infancy of this technology. Even a device as small as a ring has gone through many iterations of design and features since they were first launched in 2015. The parts are still bespoke, the tolerances are razor thin, and a huge chunk of the engineering comes down to “does this feel okay on someone’s finger all day.” It’s a weird, cool intersection of craft and tech — and it’s still very much being figured out with active advances on both the hardware and software side. Excited to see what comes next!

This week’s newsletter is a special edition written by written by Brad Hirayama of Pipeline Media Lab in collaboration with Justin Bushko, President @ Concise Engineering & Founder @ MedTech Manufacturing. Brad is a biomedical engineer and avid technical communicator. If you like this content, check out some more of his content here!

⚙️ Mechanical Engineering Resources:

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

👶 Meme Of The Week

Contract Manufacturing Basics

Most engineers approach contract manufacturer selection like a procurement exercise: gather specifications, request quotes, select the lowest bidder. Three months later, they discover their partner lacks capacity, documentation is insufficient, and actual costs bear little resemblance to the quote. Contract manufacturing is a relationship problem disguised as a procurement decision. Engineers who treat it purely as vendor selection encounter the same failures: missed timelines, cost overruns, quality issues, and eventually, expensive mid-production manufacturer switches. Success requires understanding you’re entering a partnership where both parties need specific things to work. Working with contract manufacturers requires understanding cost structures, managing technology transfer, and maintaining productive relationships. This guide provides decision frameworks and identifies common pitfalls across partner selection, pricing negotiations, documentation requirements, and ongoing relationship management.

Selecting a Contract Manufacturer

Figure 1. The Importance of Alignment in Product Development and Manufacturing Processes

Start with a minimum of five candidates. Evaluate candidates across four dimensions: technical capability, quality system alignment, capacity planning, and communication infrastructure. Engineers consistently underweight the last three. A manufacturer with perfect technical capability but misaligned quality systems generates friction on every production run. Insufficient capacity means de-prioritization when higher-volume customers demand attention. Poor communication infrastructure leaves you blind to problems until they become crises. Quality system alignment matters more than most engineers realize. If you operate under ISO 13485 and your manufacturer holds ISO 9001, expect constant translation friction. Ask specifically about their quality management system, audit history, and customer-specific requirement handling. A manufacturer accustomed to your industry’s standards moves faster and generates fewer compliance issues. Capacity planning reveals how seriously a manufacturer takes your business. Ask about other customers and equipment utilization rates. At 90% capacity, your job gets pushed whenever larger customers need something. At 40%, question why - rapid growth or customer losses. You need to know where your production fits in their priority structure.

Understanding Cost Structure

Manufacturing burden rates include direct labor, benefits, facility costs, quality oversight, equipment depreciation, inventory management, and profit margin. A $25/hour technician generates $75-95/hour in billed time. Some manufacturers separate warehousing fees (7-11% of material costs), others bundle everything. Neither is inherently better, but understand which model applies. Typical profit margins run 20-30% for standard production. Higher complexity or lower volume pushes margins higher: for assemblies like consoles and capital equipment often see 100% markups. Very high volume can compress margins to 3-5%, but these relationships carry risk. Manufacturers on razor-thin margins lack buffer for problems.

A major red flag to be aware of. Quotes significantly below market rates, unclear cost breakdowns, unwillingness to explain structure, or dramatic price changes between quotes and production. A manufacturer quoting 15% below competitors either misunderstood requirements or is underpriced to win business, planning to recapture margin through change orders. Negotiation should focus on volume commitments and payment terms rather than pure price reduction. Manufacturers respond to predictable demand. A firm commitment for 10,000 units quarterly provides planning certainty they’ll trade for better pricing.

Critical point: expect to pay for development work, samples, and tooling. Manufacturers offering “free” prototypes recover costs somewhere else like production pricing, minimum orders, or lock-in contract terms. Pay market rates for development. It clarifies the relationship and gives you negotiating position.

Technology Transfer

Documentation is heavy life prior to the transfer process: complete drawings with tolerances, bills of material with approved vendors, assembly work instructions with visual references, test procedures with acceptance criteria, and quality plans with inspection frequencies. “We’ll work with your engineering team” is not a substitute for complete documentation.

Pitfalls to be aware of: assuming documentation that works internally transfers directly to a contract manufacturer. Your internal team has tribal knowledge about material handling, common problems, and acceptable workarounds. The contract manufacturer has none of that. Every undocumented assumption becomes a production delay or quality issue.

Figure 2. Navigating the pressures of design for manufacturability and collaboration with contract manufacturers.

Ensure that you budget physical presence during technology transfer. Video calls and documentation get you 60% there. The remaining 40% requires someone on-site, watching first articles, identifying problems, and adjusting processes in real time. Plan at least one week on-site during initial production, then follow-up visits until the process stabilizes. Transfer timelines typically run 8-12 weeks from contract to first production-quality parts. Rushing creates problems that persist throughout the relationship. The manufacturer needs time to procure materials, build fixtures, train operators, and develop product knowledge. Compressing this by six weeks might save time initially, but you’ll lose it back through quality issues over the next year.

Internal Alignment and the PM Dynamic

In larger companies, the decision to work with a contract manufacturer often involves more stakeholders than just engineering. Supply chain teams typically own vendor selection and negotiations. This creates a problematic disconnect: supply chain optimizes for price, engineers deal with the consequences.

Supply chain individuals are measured on cost reduction. They negotiate aggressively, push for the lowest quotes, and prioritize landed cost above everything else. When parts arrive with quality issues, communication breaks down, or lead times slip, engineers inherit the mess. The supply chain team has already moved on to the next negotiation. This dynamic systematically undervalues the factors engineers know matter: responsive communication, technical capability, quality systems alignment, and realistic capacity planning. The problem compounds when supply chain uses RFQ processes designed for commodity purchases. They send identical specifications to ten manufacturers, evaluate responses purely on price, and select the winner without technical input. Engineers discover problems months later during technology transfer when the selected manufacturer lacks the capabilities needed or when quoted prices bear no resemblance to actual production costs.

Push for engineering involvement in manufacturer selection, even if supply chain owns the commercial relationship. Insist on technical evaluation criteria with defined weights. Make the case that upfront technical diligence prevents expensive switches later. Supply chain may control the contract, but engineering will manage the relationship. You need input on partner selection. Project managers become your primary interface with overseas manufacturers. Most contract manufacturers assign a single PM to your program who coordinates everything: engineering questions, production schedules, quality issues, and logistics. The PM translates between your requirements and their local team, converts specifications into Chinese/Vietnamese/Thai/Etc., and manages internal resources.

PM quality variance is massive. An excellent PM anticipates problems, communicates proactively, escalates early, and understands technical requirements well enough to filter noise from signal. A poor PM creates information black holes. Questions get lost, problems surface too late, and you discover manufacturing started wrong after the first shipment fails inspection. The difference isn’t subtle - it impacts every aspect of program execution. Evaluate the PM during manufacturer selection. Ask them to walk through how they handled a technical issue on another program. Request examples of how they communicate problems. Pay attention to their technical depth and English proficiency. If the PM struggles to explain their quality system or glosses over capacity constraints, you’re seeing how every future interaction will unfold. Some manufacturers rotate PMs or assign one PM to twenty programs. This guarantees communication failures. During selection, ask about PM assignment structure, their typical project load, and how long PMs stay in their roles. Stability matters. A PM who understands your product and knows your team prevents countless hours of re-explanation.

Figure 3. Key insights on engaging with your CM’s project teams.

The PM-engineer relationship often determines program success more than any other factor. Treat it accordingly. Invest time building rapport early. Share context beyond specifications. When problems arise - and they will - a strong PM relationship means faster resolution and better outcomes. Without it, every issue becomes a battle.

Managing Relationships

Weekly calls during active production are not optional. They’re where small problems get identified before becoming large ones. Establish clear protocols: who owns technical questions, who handles commercial issues, what constitutes an escalation, and response time expectations. These protocols feel bureaucratic until the first crisis. Then they prevent a three-day production halt from becoming a three-week disaster.

Forecasting accuracy directly impacts how manufacturers treat you. Forecast 5,000 units and order 4,800: you’re reliable. Forecast 5,000 and order 2,000: you’ve created planning problems. They bought material, scheduled capacity, and staffed for your forecast. When orders fall short, they remember. Next time you need expedited delivery or favorable pricing, that history matters. Contract manufacturers prioritize customers who make them money without causing problems. Profitable but high-maintenance makes you vulnerable. Another customer with comparable volume and easier management becomes more attractive. Low-volume but extremely easy builds goodwill. Low volume and high maintenance puts you first in line for de-prioritization when capacity tightens.

Making it Work

Contract manufacturing fails or succeeds based on whether both parties get what they need from the relationship. You need consistent quality, predictable costs, and reliable delivery. They need accurate forecasts, clear specifications, and reasonable profit margins. Most failures trace back to one side pursuing their requirements without understanding the other’s constraints. Engineers who succeed with contract manufacturing understand they’re managing partnerships, not transactions. They document thoroughly, communicate proactively, forecast honestly, and pay fairly. The ones who struggle optimize for initial price, assume their business is naturally attractive, and treat manufacturers as interchangeable vendors. Your manufacturing partner isn’t buying into your vision or your potential. They’re running a business that succeeds by working with customers who make them money without causing problems. Structure your relationship to be that customer.

Like this content and want more of it? Reach out to hardwareishard@gmail.com!

With all of our recent experiences in supply chain, they are extremely pertinent to the discussion of how manufacturing and procurement can and should be done.

Engineers and supply chain managers spend way too much time on sourcing. They email suppliers, wait days (if not weeks) for quotes, chase updates, manage spreadsheets, and deal with miscommunication across time zones. It’s slow, frustrating, and pulls them away from actual engineering work.

Jiga fixes that. They connect engineers directly with vetted manufacturers of CNC, 3D printed, and injection molded parts, handle quoting and communication in one place, and give full visibility into every order. What used to take weeks now takes hours.

No joke, how many emails back and forth are you getting with your suppliers? 2-3 a night max? And only if you’re working late, probably just 1 response realistically. Whether you’re bootstrapping or scaling to production volumes a platform allowing you to chat directly with them, with integrated part visualization for detailed questions and DFM is a game changer.

Intro

We believe there has never been a better time to start a manufacturing business!

Tensions with the industrial powerhouse China have never been higher (in recent history), VC have never cared so much about physical things, and AI and 3D printing give newcomers massive advantages relative to the incumbents.

Why

This website has up until now been purely dedicated to helping mechanical engineers get jobs. Coming out of school, Mechanical Engineers are some of the highest earners, especially now with the contraction in software engineering. That makes it an extremely compelling workforce to join. The median job pays $70–80k, with the high-paying entry-level jobs on the coasts offering as much as $120k+. However, a downside of the mechanical engineering profession is that there is not compounding growth in salary that you might see going into finance, business, law, or the medical field. It is crazy rare to find mechanical engineers who make more than $200k [base] a year in salary no matter their level of experience.

I think the easy way to look at this is through a quick example. Say you start at $100k and grow your salary to $200k over 20 years (3.5% per year). You would earn $2.87M and pay $1.15M in taxes (assuming you are in CA, which is reasonable given the salaries), leaving you with $1.72M in lifetime earnings. That seems quite good; however, when you factor in net present value and inflation (a 4% real discount rate is often used), that number aligns closer to $1.25M in today’s dollars of take-home.

While compared to the median earnings of college-educated Americans, this is a phenomenal return, mechanical engineers are generally a more ambitious bunch (you didn’t take those fluids and thermals classes just for fun) and this is frequently seen as a subpar return for so many years of specialized work.

To sweeten the pot there is another way to get more value: equity. While this has predominantly driven massive gains to software engineers, hardware engineers can also benefit by joining a unicorn early. The best examples are Tesla, SpaceX, and Nvidia, though given the recent surge in VC investment in hardware I am sure there will be a bunch more. The game to play here is to work at 3–7 hardware companies over those 20 years who are Series B to Series D (200–2,000 employees). You just need one or two of them to go to the moon to walk away a millionaire! The upside is clear here. The downside can be working in a very intense workplace that demands 60+ hour weeks for 2 decades while you potentially try to raise a family, get older, etc.

On the other side, one of my favorite stats is that most millionaires in America are small business owners.

This is where the opportunity for a hardware engineer to grow their wealth greater than through the typical W2 route takes shape.

It used to be that getting an MBA would be a good idea so you had a good business understanding before jumping in (or you could go back to the W2 in a management position with higher salary potential), but we would argue that ChatGPT is an omnipresent searchable MBA with personalization for your situation. Save the $200k and put it into your next endeavor!

Ok, so now when it comes to how to start/run this manufacturing business we see 3 options:

1. Raise venture to start a manufacturing business

2. Buy a manufacturing business with an SBA 7A loan

3. Start a manufacturing business from scratch(bootstrap)

1. Raise venture to start a manufacturing business

This is by far the easiest option. Change your LinkedIn profile to Stealth Startup and watch the VCs who are currently investing in all things Defense/Manufacturing (6 months ago it was crypto . . . lol . . .) pile into your DMs. Pitch a segment of manufacturing that has a $100B TAM (Total Addressable Market) and raise $2–5M seed money to start a factory. While there is the least amount of friction here, the most likely option unfortunately is not success. VCs make 100s of bets and they just need a couple to make it big and pay off all of those investments. You are making 1 bet . . . the risk is much higher to YOU. Furthermore, the best dozen funds (Sequoia, A16Z, Benchmark, etc.) see a massively disproportionate set of successful outcomes relative to everyone else and effectively prop up average returns. Those VCs in your DMs are most likely not the front of the pack. Like you, they are aiming to disrupt. This article is not about why VC-backed manufacturing companies as a cohort won’t be successful, but we think for risk-adjusted returns for any single mechanical engineer, this may not be the best return on time and effort.

2. Buy a manufacturing business with an SBA 7A loan

This option is quite interesting. There is a massive number of baby boomers who own manufacturing businesses and are retiring. Small banks will give you SBA 7A loans to buy these companies because of your degree. You have to be careful to pick the right one, but we generally think this has the highest chance of a successful outcome. The business might have 50 years of product-market fit. That list of customers, reputation, etc. is a fantastic foundation to build on. Also, MBAs who usually buy boomer businesses are scared of manufacturing because of the technical risk. That’s your bread and butter. This article is also not about this option so we won’t go too deep, but the gist of a 7A loan is that you personally guarantee it. That means if you run the business into the ground, the bank will unfortunately come for all your assets. While bankruptcy is a horrible thing, for most of our audience who is in their 20s on the coasts you probably don’t own a house or are married; therefore your assets aren’t very valuable in the scheme of things and you can just go get a W2 post-bankruptcy (note that this will impact future ability to get loans). This really confused me as there was asymmetric upside to younger borrowers. We asked a bunch of banks about it and since the loan is mostly backed by the government and they don’t want to prioritize based on assets there is no penalty to not owning a house or having a partner! Downside risk = mitigated. The upside benefit comes in what we think of as a multiple arbitrage. Imagine you grow a $5M revenue business into a $20M revenue business. Now you own a $20M revenue business (which is 4X bigger) and it will sell at a great multiple to PE (say 5–7X EBITDA instead of 2–3X). You basically can see an 8X outcome for 4X growth. One day we will dig into this further. We heavily considered going this route for a while and have a bunch of interesting data. Stay tuned. In the meantime, cherry-pick the manufacturing episodes of this podcast. It is quite dry, but so valuable.

3. Start a manufacturing business from scratch

We are basically going to talk about this option for the rest article, but we need to discuss a couple other things first.

The biggest tax loophole in America . . . QSBS . . .

The Qualified Small Business Stock exception is an incredible piece of legislation meant to help small business owners. It is even more beneficial than the classic real estate 1031 exchange (gains from your first real estate sale can go into your next real estate purchase tax-free).

We are not CPAs so we recommend you talk to your own tax advisors, but the gist is that if you buy or start a business and hold onto the stock for 5+ years, the first $30M (Trump just increased this from $10M) of gains (full sale price if you start from scratch, or increase in basis if you bought the company) is distributed tax-free on a federal level (CA still charges tax 😟 womp womp). Oh, and by the way, this is per shareholder so if you start it with 2 friends you can each do this up to $30M. Plus, as of July 2025, if you sell after year 3 it is 50% of gains or 75% of gains in year 4. There are 3 things to know:

1. You have to file an 83(b), sign it and send it in the first 30 days after creating the company.

2. You have to sell in a stock sale, not an asset sale. This means you sell the assets and liabilities, which can be more challenging than just the assets and may have fewer buyers. Granted, you may also be willing to accept a slightly lower price if you aren’t paying federal tax via a stock sale than a similar business would on asset sale.

3. You need to be a C corp for this to work. There seems to be a lot of minutiae on why someone might choose a C corp vs an S corp for their business, but we usually try to examine things from a perspective of symmetry. The way we see it, there is some upside to potentially slightly lower taxes in the short term with an S corp (disadvantages non-withstanding). They are essentially taxed as pass through entities similar to LLCs. However, the upside is completely asymmetric with C corps due to QSBS sales exceptions. Time will tell if this tactic will pay off, but our philosophy is to optimize for what moves the needle. 5% doesn’t, 50% might.

Starting a small manufacturing business from scratch!

Okay, here we go. Buckle up and get ready.

We are kicking off a new era of Hardware is Hard focused on starting a manufacturing business in America, because we quit our jobs 6 months ago to do this and it is fresh on our minds.

Starting a business from scratch, especially one you bankroll, is exceptionally risky. Hardware takes time to build and even longer to reach product-market fit. One of my favorite non-VC hardware company start from scratch is Opulo; they have grown their pick-and-place business from a garage to a massive entity and documented it along the way. Check it out!

This is why we believe that starting from scratch for a product is a bad idea. Or, perhaps not a bad idea, but a risky one. Instead, go into a business where product market is almost guaranteed. As we mentioned earlier there is a ridiculous amount of VC money pouring into hardware companies. Most of them are building hardware products (even though they want you to believe they are building AI software products…). The American industrial supply base can’t move as fast as these companies need. The great subcomponent manufacturers (maybe even the one you buy with a 7A loan) have been putting away cash for decades making the same things since the 80s. They are extremely averse to growth or risk, especially since they are about to retire. Why open a new facility to meet higher demand when you just want to work 3 days a week and go sailing on your boat? Find a startup you believe is gonna win (under 100 employees). Find a supply chain gap, and go all in delivering that thing as quickly as possible. Sell yourself as a product and a service (you are an engineer). Why target small, quickly growing companies? Rising tides lift all boats, and if you can supply them, as they gain traction you will to. Also, it is WAY easier to get in the door and be taken seriously at smaller companies. A bigger company may have layers of defenses (bureaucrats) and it will not only take longer, but potentially be unjustifiable for them to trust you with a critical business need. Their risk tolerance is much lower than smaller companies. Another layer of defense is that you may actually be able to physically show up on-site randomly and meet with someone at a smaller company who will discuss their problem with you. Walking in the front door does wonders relative to DMing.

For the product vertical that you choose, it is also important to try to stay away from research projects. Research projects (i.e. not a solved problem) are great for defensibility, but have indefinite timeline and require undefined resource input. Stick to things that are solved problems, where integration can be tricky but not impossible. In this case, the goal is not to invent something completely novel (at least at first). It’s not as sexy, but the goal is to make something in the US, quickly and cost effectively, and to reach that goal, you cannot be putting all your chips on a science project.

Jumping In:

Depending on your risk appetite, there is a certain level of maturity you may desire before jumping in. That’s okay, you can do some stuff on your nights and weekends, but in this game, moving fast is big. If what you found is an underserved market and you are serious about starting a business, you will have no choice but to jump in and get moving. Any time you spend working nights and weekends is time that someone else is spending working days, nights and weekends.

That said, here are a few things it might be good to consider fleshing out prior to pulling the trigger:

• Market evaluation / Location

  • Look at growing markets, find the most likely winners (look who is backed by the big VCs), and talk to them. Find their pain points, and see if there’s somewhere you can help address it. Move your location to wherever they are. Proximity is a powerful plus, and no one else is likely to pick up their existing business and do the same thing. You’ll be starting from scratch, and they will love working with you if they can see and trust you in person.

• Technical feasibility

  • Make sure that whatever pain point you’re trying to address is something you can address, ideally with relatively low capital expenditure (at least at first), and ideally relatively quickly. I personally wouldn’t decide Silicon fabs are growing fast, then look at ASML/Intel and try to make ultralow roughness mirrors or lithography masks. Those are things which are extremely difficult to make and cost millions per unit. Start with something that seems more expensive than you would think it should be, but still has an individual part cost significantly less than whatever capital you are injecting into the business.

• Work life balance

  • Be prepared not to have it, at least for a while. If you want to go this route, that’s just what it takes. It’s okay if it’s not worth it you, but be realistic with yourself or you’ll end up miserable.

• Cofounders

  • Trust is everything. A lot of people decide they want to start a company then go looking for cofounders with similar vision. Our experience is a bit different since we’ve been best friends for 7 years, lived together for 4 of those, and have also been working on Hardware is Hard together for the last 4 years. We’ve done a ton of projects together and this has always been a dream of ours. Alignment is absolutely vital here so making sure that you two can have vehement disagreements while still maintaining complete respect for one another is a must. The disagreements will come. If you’re both all in, this will probably be the thing you both care about more than anything else in the entire world. Both of you will want success to occur, but you may have different ideas on how to get there. Clear and consistent communication is everything.

In terms of actually pulling the trigger on quitting your job, I would recommend thinking big picture here. “Oh but my stock vests in two months” or “promo cycle is in November” will ALWAYS be true. I’m not joking, my previous company had stock vests in March, April, September, and October, perf review and EOY bonus in November, and ESPP purchases in February and August. That’s intentional. Makes it really hard to quit when you have perceived “big chunks” coming in. Remember the asymmetric upside and be strong in chasing what you want. Note that I do recommend getting out of any student debt and squirreling away enough money to survive (read: survive, not thrive) before doing any of this.

There’s rarely a “great” time to quit your job, but how you do it can have real impacts on how people remember you. Be respectful but firm, give 2 weeks+ of notice (but depending on industry, be prepared to be walked out in the worst case same day), thank your boss and coworkers for all you’ve learned, and be honest. Honest and direct feedback is the best way that organizations can improve. Don’t twist any knives, don’t badmouth, but do tell your boss what could have positively impacted your time at the company. My personal opinion is that you should tell anything you have to say to your boss/their boss/coworkers. Telling HR is essentially a complete waste of time; maybe that’s a hot take but my experiences have affirmed that statement in every single possible encounter. They will schedule an exit interview with you. The important ones will be with your own management team. I recommend preparing ahead of time the things you think it’s important that they know, and if they don’t schedule an exit interview with you, schedule it with them. You have nothing to lose and they have everything to gain. Exit with tact and say your goodbyes fondly, those connections may become more important than you think at anytime down the road.

Starting the actual business:

This is a really brief summary of the highlights. Over the next couple of years we will go into each part in depth. For now here are some of the big gotchas:

The first thing you need to do is build the product. Do this in the most capital-light way possible, though it should be as close to production representative as possible. Don’t rent space, use a garage. Don’t buy machines, outsource the parts. 3D print everything to test. Borrow, be creative, and think scrappy. Build that first product in weeks and deliver it to your potential customer as quickly as possible (honestly whether or not they ask for it). Show up, say you have a delivery for an engineer you friended on LinkedIn. This is gonna be the first of many, many sales endeavors.

Once things are going well with the company and it looks like you might secure an order (of any size), form the company. Use Stripe Atlas (lawyers you can afford will suck and cost you time). You need to be a C corp (remember QSBS). Loan the company some money ($10k–50k) and set up QuickBooks. Tracking finances matter much more when you own the company instead of work at it (as unfortunate as that sounds). We will dive further into the finances in a couple articles later, but for now, we recommend starting early and reach out to friends/family/professional network for a good CPA. Talking with them early can save a lot of time (and money).

Supply Chain / Quality

When you own the business and you aren’t venture-funded, the incentives around supply chain completely change. What used to be about moving fast (Protolabs, Xometry, OSH Cut, Fabworks)

turns into saving on cost (margins really matter) and guaranteeing quality, transparency, and timeline. Your quality is your reputation. It is what distinguishes you from your competitors and justifies your higher prices than Mexico or China (even with tariffs).

This is where we want to talk about the company partnering with us to tell this whole story, Jiga.

Jiga is a sourcing collaboration platform with a carefully curated network of vetted, high-quality suppliers offering ITAR, AS9100, or ISO compliance. And the best thing is, they can connect you with American companies in addition to foreign suppliers. They don’t hide who you are working with. Time and time again when I used Xometry for multiple orders, the parts in the second order would be different from the first one because they swapped suppliers on the back end without telling me. Nothing erodes supplier trust like obfuscation. Jiga gives you full transparency into your supply chain, and when things get complex, there’s a real team behind the platform that helps you navigate it. You’re not submitting a ticket into the void. You’re talking to people who actually understand manufacturing and can help you make the right call. We are going to do a huge market study on Jiga and their competitors (get ready for financial analyst-level digging with an engineering-first approach), but in the meantime, try out their product. We are already using them and are huge fans.

We don’t get credit for sending you unless you use our link jiga.io/hwh. We know you are probably reading this on your personal email/laptop. Please go to jiga.io/hwh when you get into work. We only take on a couple sponsors and truly appreciate you using our links!

Ok, a couple other quick things:

Chat GPT (Your MBA Wingman)

Leverage ChatGPT as much as possible. Create projects for Finances, Compliance, Corporate Structure, etc. so it remembers your history. Use it to figure out how to start a business bank account, get sales tax exemption, hire people, etc. Pay for the plus account. It is worth it 10X over. Oftentimes when we talk to people about using ChatGPT, they say “it hallucinates,” “it gets things wrong,” “it takes too long.”

A. You might be doing it wrong. It needs as much context on you and what you’re doing as possible to be good; provide all that context.

B. It being 80% at anything might seem horrible when you are a subject matter expert (aka in engineering), but one thing we have learned is all those boomer businesses and any other small business is already doing less than 80% at things like law, compliance, and finance. We have been shocked how much better it is than when we pay for those services. And . . . most of the world isn’t 90+% compliant, etc. Welcome to small business. So move fast and leverage ChatGPT. It is one of your biggest advantages as a young small business owner. (Don’t have it do your finances or payroll though; trust and pay for QuickBooks). Move fast and break things!

C. Iterate. One-shotting is a relatively rare event. Sometimes it may take 5-10+ prompts to get desired action on a complex script. Is it frustrating? Yes. Is it 1000X faster than you writing the script from scratch? Yes. I know it takes some of the fun out of engineering, but the 10X speed is worth it.

D. Be wary of rabbit holes. If something seems harder or more complex than it should be, it probably is! Using ChatGPT is not an excuse to turn off your brain. If something doesn’t make sense to you, find out why. The number of simple mistakes you can still find in LLMs is shocking, they are definitely not at the stage where blind trust will work. That said, I have gotten to the stage where if something sounds right to me, I will usually accept it without further proof. Perhaps this means I am creating an echo chamber, but generally speaking first principles can get you pretty far.

Real Estate

You aren’t venture-backed, so you don’t need beautiful offices. Find space on LoopNet that costs $1.25–1.75/sqft and only rent what you need for next 6 months (1–3k feet). If you are spending more than 2X your apartment’s rent, you are renting too much space (or have too many roommates). Expanding is a good thing (you can then afford more space!). Who can. say what your needs will be six months from now, there is no data to extrapolate based off of.

Also, don’t sign a contract for more than 2 years. Brokers will give you an AIR lease. We learned the hard way you basically can’t change it. Maybe if you use a broker of your own and have them negotiate on your behalf, but be prepared for that to waste a TON of time. Instead, focus on any changes the landlord made to the standard boilerplate (typically in red) and any addendums (where they talk parking, signage, noise, etc.). You want a gross lease, not a triple net (NNN) lease (ChatGPT the difference if curious) and we recommend you visit a bunch of places. Shotgun approach here is crucial. You will not know the market or if one place you tour is a good deal or a bad deal unless you see 10+ data points. We were shocked with the wide range of options for the same price (as well as by the responsiveness and availability of leasing agents). The variance was much higher than an apartment. Also, renting from a building that is owned by a person or two is much easier than one owned by a REIT. Ask about this when you visit.

Finally the last thing we learned the hard way is look into the power. It is one thing to have the breaker capacity; it is another thing to have the actual wire run to where it needs to go. If you are going into a product that requires big machines, make sure the power runs exist. It is incredibly expensive to install those later (think potentially a month of rent+).

One underrated item that sometimes gets overlooked: parking. How many employees do you envision having six months from now? Where will they park? This often isn’t a dealbreaker, but it definitely something to watch for.

Also note that the commercial real estate market is cold right now. Luckily, that’s great for you! That means places sitting on the market for months —> more supply, less demand —> desperation —> better prices! For now, YOU are in the drivers seat. Don’t accept getting hardballed and know that negotiations are good! The worst they can say is no.

An actual photo of our warehouse the day we moved in! Though unorthodox, those red floors have really grown on us (…or at least one of us…).

Machines

Don’t get greedy. It is incredible what people can do. Don’t automate to start; your product will almost certainly change a bunch before you lock in big orders. Use Facebook Marketplace. It is way better than Craigslist these days. Chinese machines are also a possibility. There is a lead time and functionality/service risk, but it might be worth it for a quarter the price (or even less). Often the big American manufacturing companies won’t even sell to you yet because you aren’t big enough. On the other hand, it might make sense to buy the American brand if the machine is the core of the product you are making. Service is highly underrated, being able to talk on the phone to someone that understands what you’re saying without translation and can help fix issues that come up is BIG. This is more and more true the more expensive the machine.

A friend of ours postulated that of all the equipment he’s bought (a lot, trust us), 90% of the time, the amount of time he spends debugging the Chinese machine ends up being about equivalent to if not higher than the extra money that he would have spent on the American machine. Depending on the machine complexity, that’s totally a valid point. For simpler machines, your basic engineering skills should be enough to reasonably quickly fix most problems that crop up.

One of our facebook marketplace finds is a China classic and has broken 3 times in the last two months. However, fixing each break involved:

1. Recrimping a wired connection correctly

2. Replacing a terminal block

3. Replacing a relay

Root causing can sometimes be a difficult and has no exact timeline, but for simple machines, there’s really only a few things that can go wrong, so an approach of isolate components and check if there’s an issue usually works.

Also note that single point of failure = bad. If it can fail and bring your line to a halt, it will. Never underestimate the power of redundancy. By redundancy I mean stocking extras of components after they fail once, not necessarily stocking an entire extra machine… unless it’s justified — we’ve done this with our most critical machines. Having a way to do a method manually as a backup for machine failure is perfect until it goes back online, but if there is no way to manually replicate it, a complete backup might be a good idea (plus if it’s a line bottleneck, then this will expand production — downside is that it’s probably costly).

People

The job market for manual labor is super tough right now, so there is a ton of well-qualified manufacturing labor out there. Spend $45 on a Craigslist ad and then invite the top candidates in for an actual test trial (pay them to actually do the job for a few hours). Pick the best!

Ok, we’ll leave it there for now. Stay tuned for more. Get the wheels turning on what opportunities might exist. Good luck!

If you made it to the bottom and are interested in more content like this please email us!

This week’s newsletter is a special edition written by Madhav Bhat based on interviews with several AI Text-to-CAD founders he spoke with in preparation for our last article! Madhav graduated from the University of Michigan with his BSE and MSE May of 2024. Throughout his recruiting journey he interviewed for mechanical engineering internships and full-time positions with companies from large Fortune 500 corporations (think Ford, Michelin, Apple) to growth-stage companies (think Tesla, SpaceX, Anduril) to startups ranging from 3-1000 people in size working on all sorts of projects. He is interested in writing more content on the wearables space, so if you or someone you know works in wearable tech, PLEASE reach out to hardwareishard@gmail.com, we would LOVE to get your thoughts.

⚙️ Mechanical Engineering Resources:

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

👶 Meme Of The Week

🗣️ AI Founder Interviews

In addition to using the tools, we also had the opportunity to chat with founders of 3 AI CAD tools/startups: Zach from Adam, Cameron from Ballista Labs, and Jordan from Zoo on the state of AI CAD tools and the vision for their products as well as the industry as well as a bunch of other interesting topics. Here is a synopsys of our Q&A session:

Q: Why did you start your company?

Zach: We were working on Generative 3D research at Berkeley at the time LLMs were taking off. LLMs were very good at writing creative software for people with no software skill, and people were really struggling with CAD.

Cameron: We pivoted to get here. We had an “aha” moment with Gemini 2.5 that it might be a better engineer than I am. Ask it to design a heat exchanger and it might design it better than I could. From there - we had the idea of what if we orchestrated 100+ LLM calls for a massive engineering project. Can LLM calls perform the work of a 10 person engineering team? The problem there is it can design systems but things would start breaking down when it came to geometric considerations. Therefore, we needed a CAD module to solve this problem.

Jordan: My cofounders had experience in the software space and I had founded Relativity Space. They saw what we were doing on the factory floor and said “they must be doing something different in software” and our answer wasn’t great. We put out a call to action for someone to take on this space and realized we could do it ourselves.

Q: Who is the ideal customer of your tool?

Zach: We analyzed during YC “what is the state of AI CAD” and found this tech is very early. What it is helpful for is if you quickly have an idea and want to create a model to 3D print. We’re going more consumer to start and using that as a base, up to mechanical engineers a year later. Now our audience has shifted and we’re building for professional mechanical engineers. We’re currently building an AI agent that works inside of Onshape.

Cameron: Primarily people who might not know CAD or simply want to work/iterate faster.

Jordan: Our view was that if you introduce proper AI CAD, you need to go full stack and hybrid. By a hybrid approach, we mean combining both text-to-CAD and traditional mechanical engineering workflows to cater to both consumer and enterprise.

Q: What is the goal/vision for your product/company?

Zach: Our original mission of the company: with AI, professional knowledge work and creation knowledge work will change a lot especially in the engineering disciplines. We can orchestrate agents that do the work for us. It’ll be more about making higher level decisions. Design intent will still stay but a lot of the work will be handing off to agents.
Cameron: We are heavily influenced by the progression of LLMs. By the time Gemini 4 comes out, people will be shook by how good it is. Capabilities of LLMs will supercharge what Orville can create and I’m a big believer in test of time. Right now it can do one complex part or an assembly of 3-4 complex parts. It’ll go to 10-20 parts and 8-9 months it’ll be genuinely useful for engineers.

Jordan: Our goal is a full platform end to end hardware development. Design, manufacturing, analysis, etc. For example if you have an anomaly failure at Boeing, you need every piece of data analyzed. You need traceability. What better way than have a full development platform that does that. Engineering is so much more than design and the best copilot will have the best synergized dataset. However, the Core IP is geometry. Our view is there are a ton of MechE software companies doing AI requirements tracking, supply chain, etc. That is all influenced by geometry…

Stop. Before you automate that manual process step-by-step, read this.

In the 1890s, inventors tried to ease the transition from horses to automobiles by building “horseless carriages” - mechanical contraptions with articulated metal legs, rein-based steering, and fake horse heads housing lamps. They were solving the wrong problem. The answer wasn’t replicating a horse. It was rethinking transportation entirely.

Modern automation projects make the same mistake. Your manual process evolved around human capabilities. Operators compensate for variation, use visual judgment, make micro-adjustments. Automating these workarounds directly creates complex, unreliable systems.

Pipeline Design & Engineering’s focus? Solving for the optimal outcome. Clear, reliable, effective.

Don’t automate the workaround. Solve the actual problem.

Evaluating automation? Pipeline has solved this puzzle 100+ times.

Contact Us

Q: Do you foresee problems with enterprise adoption and getting legacy companies to adopt these tools/do you see yourselves competing with a Solidworks/CATIA/NX?

Zach: SWEs are excited to try latest technology, MechE not as much. That’s a big reason we started with Onshape since the users are keen to try using ADAM. They want to use the most modern tools out there. It’s going to take more time to build this than software engineering agents. Cursor was able to fork VScode. We don’t have that luxury. The most we can fork is freeCAD. It’s too early to say we’ll be a competitor of Solidworks. PTC has been great. We’re big fans of Onshape.

Cameron: This will be a new way of operating. It’s not a parallel with coding tools, you’ll have CATIA and Solidworks, but the user will always be at the top of the pyramid of work and Orville will solve everything underneath.

Jordan: There is a misbelief in the market that CAD will never change. The market was set up for disappointment for customers by companies that never offered anything new. The functionality hasn’t changed in 30/40 years and platform agnostic, CAD has had the same workflows. If you rebuild the stack which is a huge investment of time and labor, the tools will move the market. That’s where the market bifurcates. People go add-on, or separate workflow, and that feels low conviction. This is starting to get proven out on our side. We’re working with customers such as subsidiaries or specific teams on one project and then snowball through the company. A year ago we did not have that momentum.

Company Specific Questions:

Q for Adam: Talking about the Onshape integration (one of our sponsors!), will you all soon build out integrations for other CAD tools?

Zach: We are open to building on other tools. Transparently, people have heard nightmares with people working with Dassault (Solidworks maker). We are open-minded. The main thing is figuring out how fast we can get this tech to the hands of people. No one has built the final iteration of this yet.

Q for Orville: What are some insider tips/tricks to maximize Orville’s usefulness?

Cameron: The biggest thing is embracing the ability to do follow-up chats. While we can build things in one shot, you can get great outcomes by staying for 2-3 more chats. It’s similar themes from prompt engineering - bring everything you want at the highest level. Don’t tell it how to do it and the exact steps to take, keep things high level and it’ll do a really good job.

Q for Zoo: I want to talk about some of the more interesting design behind Zoo, including the code (KCL) backing, hybrid approach and GPU acceleration. Why make these decisions?

Jordan:

• For KCL: this helps bridge CAD to a large language world. Words represent geometry rather than tokenizing meshes or STL.

• For hybrid: if you don’t have hybrid you can do simple changes, but you can’t add a feature. You can’t change how they are related to each other. How is a hole sized a certain way, how is it toleranced, etc. That’s the magic that makes mechanical design work. AI trained to augment mechanical engineers have to have that exposed in training. Would people adopt cursor if it just generated compiled code rather than code you can edit?

• This ties into reinforcement learning. We launched Zoo Enterprise where our primary product is fine tuning text-to-CAD on enterprise data. Customers have millions of CAD files and we can do reinforcement learning on those. Do a generation and see what a user edits and what is the downstream change. In production that context exists. We can fine tune on their data and lots of their metadata and get info from tickets, assemblies, etc. Not everything is written in an obvious spot. Integrate with more systems and get more satisfaction.

• On GPU acceleration: the winner in the space will have the best compute stack. Long term this is a compute game and ideally we would have the full stack integrated by us to train models faster, cheaper better. This won’t be seen until the long term.

Conclusion

Ending with some shoutouts:

• If you are interested in beta testing Adam’s Onshape integration, reach out to zach@adamcad.com !

• If you want to try out Orville head over to ballistalabs.ai!

• If you are an enterprise customer looking to advance your CAD workflows using AI head over to https://zoo.dev/enterprise!

Once again I was surprised by how good these products have gotten but talking with founders I realize we are still SO early. Enterprise adoption has only just begun and it’ll take some time. This is only a cut version of our full interviews with these founders and it was eye-opening to see how many approaches and visions there are to solving AI CAD. Excited to continue watching this space!

🦸 Guest Writers: Madhav & Sajiv!

This week’s newsletter is a special edition written by Madhav Bhat and Sajiv Shah! Madhav and Sajiv are both from University of Michigan’s Mechanical Engineering program, and love hardware just as much as we do. I personally found the survey on AI CAD tools very interesting and can’t wait to try it for myself.

⚙️ Mechanical Engineering Resources:

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it take to be The 10X Intern

💻 Exploring AI tools for mechanical engineers

When ChatGPT first dropped I was in my junior year of college and it totally changed how I learn about all kinds of subjects (including mechanical engineering). Now in the software world, tools like Cursor and Claude Code have become staples in the workflows of software engineers and it begs the question: how will tools evolve to meet the needs of other engineers, especially mechanical? To try and learn more about this I’ve been trying new AI-native tools for mechanical engineers, specifically for mechanical design. I first started with web and text-based design tools and the main ones I came across were Adam and Zoo. Both of these have a text-to-design mode as the first-class entrypoint to design but different approaches following that.

Adam:

Adam was started by Aaron Li and Zach Dive when they were at Berkeley’s Master of Design program. They recently raised a $4.1M seed round to power the future of CAD.

Prompts I tried:

For my first prompt, I was a little too excited and asked it to model something complex: “Model a hexagonal nut with chamfered edges and standard M8 thread dimensions.”. The response I got is:

“Since you’re looking to create a hexagonal nut with specific M8 thread dimensions and measurements, you should try the **parametric mode** to get the best results! Parametric CAD is much better suited for creating precise hardware parts with standard dimensions and threading specifications. The mesh modeling tools I have access to work better for organic shapes and artistic creations rather than precision-engineered parts like threaded fasteners.”

I had difficulty finding what parametric mode referred to. It gave me this answer whenever I asked for models with parameters (like “Design a simple box with an open top and 3 mm wall thickness”). I decided to try out some simpler prompts like: “Model a flat plate with rounded corners and four evenly spaced holes near the edges” which gave some better results:

Adam UI with chat interface, visualizer and export options

I reached out to Adam founder, Zach Dive, to find out more about the capability of the tool and he pointed me to their open-source repo here. This was more what I was looking for in terms of AI CAD. The parametric mode allowed me to query for more complex objects like ““Make a solid cylinder with a diameter of 40 mm and height of 60 mm” and it created the part with sliders on the right-hand side to control the parameters.

To try something more practical, I provided the prompt “Design an ISO-standard M10 bolt with threads modeled” and then refined the prompt afterwards to ask to be able to modify the height of the bolt head. This also worked as expected. I didn’t download the STL file to validate the result, but the parameters on the side aligned with what is expected for an M10 bolt.

The team is also working on an Onshape integration! Try out Onshape here if you haven’t already!

ZooCAD:

The second tool I tried was ZooCAD. Zoo was founded in 2021 by Jessie Frazelle (former founder of Oxide Computer Company), Jenna Bryant (partner at Embedded Ventures), and Jordan Noone (former founder of Relativity Space and Embedded Ventures).

Their blog post does a great job highlighting the capabilities and vision for Zoo as a brand new tool with a new geometric engine for modern CAD.

ZooCAD was a bit more overwhelming at first and looked more like a typical CAD editor with an interface at the top to adjust your designs after the 3D generation. I started with the tutorial which builds a 9-blade CPU fan with its housing. The first thing I told ZooCAD to do was adjust the 9 blades down to 6. It took 30 seconds to reason and then popped out with this model.

ZooCAD implies some CAD familiarity with the bar at the top with typical CAD features, but overall the interface was simple to use and I was pleased with its prompt understanding.

This tool seemed a bit more direct of a Solidworks alternative as opposed to a secondary tool.

I also tried some more complex queries here like “create a new design for a casing for a raspberry pi 5 with a 2.5” extension for a keychain loop”. The tool reasoned for about 5 minutes before outputting the part below, which wasn’t exactly what I was looking for but decent.

Opening up the reasoning it did mention it had a 0.64 confidence rating that the output part aligned with my prompt and that there were some discrepancies between the output and the description, so I tried the same prompt as Adam: an M10 bolt with threads modeled. It reasoned for ~10 minutes before outputting the model below.

This also had a similar confidence score in the reasoning so I was skeptical if the dimensions were correct. Looking at the code editor it pulled the correct major diameter and thread pitch, but I didn’t validate through downloading the STL.

One thing super interesting here is the code editor, which uses KCL (KittyCAD Language) which looks to be proprietary to Zoo, but represents the CAD model in a fairly human-readable form. Zoo highlights the main value-adds here are being able to parametrize constants (which allow you to integrate complex assemblies and integrate parts easier) and reuse components. I’m very interested in exploring this more.

Overall I was pretty pleased with Zoo as well and liked that I could use the traditional CAD surface tools in addition to both code, and the text-interface. It’s worth mentioning that Zoo has both a desktop app and a web platform and I only tried the web platform.

👶 Meme Of The Week

A very mechanical engineer Christmas tree!

🙋‍♂️ Interview Practice Question of the Week

A steel tube is placed inside a hollow aluminum tube of equal cross section area, and equal length. A heavy plate is then placed on one end of the tubes. What is the load distribution between the the two tubes?

✅ The Answer

This is a common question asked by Tesla and SpaceX in early interview rounds. Fundamentally, employers are looking to evaluate your understanding of statics and how you can apply equations. When the question asks for the “load distribution” this should clue you in that you are looking for some ratio of the forces that each tube is reacting from the load.

The key to solving this problem is realizing that under load, the tubes will have equal strain. Using the definition of stress/strain and how they relate to force and area for a tube in tensile/compressive loading, we can set up an equation that relate the forces to the area and young’s modulus of each tube (known variables), and ultimately come up with a formula for the load distribution.

Since the area of the tubes is equal, the load distribution is solely dependent on the stiffness (Young’s Modulus) of the two numbers. It is useful to know for interviews that steel has roughly 3x the stiffness of aluminum!

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it take to be The 10X Intern

👶 Meme Of The Week

🙋‍♂️ FEA Field Guide

As I began my journey in mechanical engineering, I remember vaguely knowing that FEA was something that I needed to know how to do, but not knowing where to start (iirc the list was FEA, Arduino, Matlab and LabView - I guess those are the skills I thought of as related to engineering when I was 18). It’s a daunting challenge to begin on your own - what software to use? What material considerations do I need to take into account? What type of analysis fits my load cases? What’s a contact set? And how is a fixture different from that? Mesh size? These are a few of the basic questions I’ll try to address below:

FEA in pretty much any software has a general flow to it:

(Hand-calcs) —> Simulation Type —> Geometry and Material —> Fixturing & Contact Sets —> Loading —> Meshing —> Results

We had an avid user submit a design they had for some at-home gym equipment made out of sheet metal! This is one of our favorite ways to come up with a personal project: need something, find it on Amazon, decide it’s stupidly expensive ($189 in this case), then design and fabricate it! Their project is what we will be doing practice analysis on for this article (read till the end to find out how much it cost).

[ 0. Hand Calcs ]

Generally speaking, hand calcs are invaluable. They are a great way to provide a quick temperature check prior to running simulation (which can take time to setup, mesh and run). That is what ~50% of our newsletters have focused on historically: How to simplify real world use cases down to a basic load case on a simple enough structure that we can estimate the stress or deflection in (coincidentally those are the most common interview questions). It’s an invaluable skill that will honestly change how you view the world over time. However, when working on something where every gram of mass costs $$$ or where “probably fine” just isn’t good enough, you’ll probably have to bring out the big guns (FEA).

1. Simulation Type Selection

Picking the right type of simulation is like picking the right lens for a camera—each type lets you see the problem in a different way. Usually there will be one way which you are chiefly interested in (dominant failure mode). Your software will typically have options like static structural, modal, thermal, nonlinear, or dynamic simulations, among others. The choice boils down to what you’re analyzing and how it behaves under loads. Here’s a quick breakdown:

• Static Structural: TBH this is the one you’re going to use 90% of the time unless you specialize in FEA. Static structural is for cases where loads can be approximated as constant with respect to time. This is the bread and butter for most mechanical designs, and a crucial part of most complete design processes in industry. For reference, imagine a spring finger on a plastic injection molded piece you’re designing. Static structural analysis can help you figure out how much deformation it can sustain before yielding, where it’ll yield, and maybe even calculate insertion force.

• Modal Analysis: As the name implies, this type of analysis is mainly for finding the shape of the primary, secondary and tertiary modes of the component as well as the natural frequency. This is very popular for aerospace and automotive particularly, as large components under a steady vibration at the wrong frequency can undergo tremendous force (and deflect wildly).

• Thermal Analysis: This is pretty self explanatory. If you need to assess how temperature affects materials (useful for electronics or high-heat applications) choose this one.

• Dynamic Analysis: This is for situations where loads or constraints change over time, like drop or shock impacts. As you might imagine, this type of analysis (and testing) is very popular in the consumer electronics industry.

Try it in Onshape!

2. Geometry Import & Material Assignment

Most software allows you to import a CAD model, but it’s essential to review your model before jumping in. Simplify wherever possible. Remove unnecessary details like threads, embosses, or bolt heads that don’t influence the structural integrity (every tiny feature adds to computation time). Intersecting geometry or broken surfaces may also result in a non-convergent model.

Many simulation softwares have a built-in material bank. Personally, I tend to use it and cross reference a few critical parameters with Google but leave the rest alone: Young’s modulus, Poisson’s ratio, density and yield strength. CTE if it’s a thermal simulation. Many larger companies will have an in-house materials database finely tuned with real world test data relevant to them. Also of note is that sometimes properties like CTE are non-linear across temperature, so make sure to input values at the correct temperature or use a table of data containing parameters across temperature (though you’ll still have to specify temperature later if it’s relevant).

The golden rule (of FEA but also data and maybe life(?) in general): garbage in, garbage out - select the material that aligns with the real world scenario you’re trying to simulate. If some properties are estimates, that’s okay, but make sure to call it out as an assumption and potential source of inaccuracy when presenting the results. I cannot stress this enough, it’s basically a motto at this point.

3. Fixturing & Contact Sets

Fixtures define the constraints—essentially, you’re telling the model what parts of it are held still or locked down, so it doesn’t start spinning off into the ether when a load is applied. In FEA lingo, this is called a boundary condition.

Fixturing can be as simple as saying “this face is fixed” (as in the case of a Command hook adhered to a wall) or more complex, where certain areas allow movement only in specific directions (a pin in a slot). I’m no FEA guru, but typically I fix a face or body as the starting point of my analysis, then go through how everything else interacts with it. For this analysis, I’ve chosen to fix the central post.

When working with assemblies, contact sets tell the software how parts interact—whether they’re bonded together, sliding against each other, or can separate under load. Different types of contacts, such as bonded, sliding, and frictional, simulate real-world interfaces between parts. Note that there are many different names and types for these contacts depending on the simulation software you are using. Onshape, for example, appears to use “mates” to dictate part interactions.

Try Onshape

4. Load Cases

Load cases are where you define what forces, pressures, or thermal conditions your part will experience. This might be as simple as a static force applied to one face, or it could involve a stepped loading condition or enforced displacement.

Careful definition of load cases is essential to getting useful results. Luckily, it’s also relatively self explanatory most of the time. For example, here I’ll apply a load approximating weights pulling down on the weight bar, and someone lifting from the cable attachment bar. For this simulation, that means 200 lbf in the -Z and +Z directions on the appropriate bodies.

5. Meshing

In short, meshing divides your model into small elements that the software can do calculations on. I won’t get into the minutia of it here, but the basics of how FEA works (i.e. the finite element method - FEM), is by breaking up part geometry into thousands of tiny triangles (depending on the mesh type you’ve chosen) and modeling the interactions between each triangle. As each triangle only experiences a slightly different stress/deformation relative to the triangle next to it, we are able to approximate the entire model. This is, of course why ruptures and cracks are very difficult to model, though there are some cool meshing features coming out that can help with that these days (basically remeshes everytime a deformation threshold is reached).

In principle, the finer the mesh, the higher the accuracy, but this also means longer computational times. Also, garbage in, garbage out. A fine mesh will not make your model more accurate if it’s setup improperly. Plus, there is absolutely a point at which further meshing functionally does nothing to improve your model. Mesh where needed. Resist the temptation to apply a super fine mesh everywhere.

First pass, I usually apply an adaptive mesh (basically “auto-mesh” based on the software’s built-in capabilities). I then inspect and apply a finer mesh to any faces of particular concern (expected stress concentrations) or small features that haven’t been scrubbed out in the initial geometry preparation. By doing this, I am attempting to ensure that areas of high gradient (high changes in stress or deformation per unit length) are well described by the mesh to help with convergence.

General rule of thumb is to have at least two elements across any surface, and more elements in areas of high stress gradients. Small holes can add a shocking amount of computational complexity (mesh has to weave around and into the hole). Fillets might seem like a great thing to take out, and they can be, but not in areas of high stress. Corners or point contacts can easily create what is called a “singularity.” Basically, it is a region where load is divided by zero (or some very small number) and stress is calculated as infinite. This is frequently the source of divergence when a simulation is taking longer than you think it should to run.

In Onshape, there basically are no meshes! Fortunately, I found Onshape’s simulation capabilities to be very generous in terms of convergence. Convergence means that the software thinks it has found a passable solution. This cannot occur if parts are flying away (if unconstrained properly) or if the stress far exceeds the ultimate strength of the material. Onshape has a cool feature where it displays iteration results even while it works towards a convergent solution, so you can literally watch the results refine as the model moves toward ultimate covergance. The upside, is that you will almost always be able to get a model that has results relatively easy with a proper setup. On the downside, it becomes much more difficult to refine the mesh or remove point discontinuities for further tuning/accuracy. Luckily, if that further accuracy is required, you can still import the mesh from elsewhere as desired and run the analysis like normal.

6. Results

Results time! Looking at the .gif, I don’t see much in the way of obvious issues (note that deformation is scaled 100X). The yield strength of 304 stainless steel is ~205 MPa, (why I set the far right indicator of the legend at 205). The fact that the max stress of the model is 3331 MPa means that there is definitely a singularity. Looking closely, it seems that the plates experience almost no stress, though the cable bolt is a bit of a hotspot (though well below yield). It does however present itself as a weak point, especially since it is currently in double shear (only very low clamping force can realistically be achieved in this part assembly). For that reason I would recommend adding in a spacer (to enable bolt clamping) and potentially a thicker bolt. The plates can be pocketed or thinned substantially as they are extremely overbuilt for this scenario. Then again, gym equipment can experience a lot of unexpected forces (dropping heavy weights) so the downside to overbuilding the plate isn’t really a downside at all.

Zooming in on the end of the bolt, we see some odd streaks of very high stress. When encountering just a couple elements showing shockingly high stress, they can often be ignored. Those stresses just aren’t real, but rather vestigial artifact of the mathematical model used to run the simulation. They can sometimes be removed by going back and looking at how the part was constrained in the first place and may lead to further improvements in the accuracy of the model. Here, it is likely due to Hertzian contact (single point rounded contact surface) causing stress singularity. Shear or deformation of the cable bolt is likely the dominant failure mode for this system, however these stresses actually appear closer to the ends of the bolt rather than the interior edge of the plate (where shear would occur). As such, they can be ignored for the moment (or further refinement of the contact set could also probably solve the issue).

Try It Yourself!

Et voila! The final result speaks for itself! I am told it works like a charm! Total cost of parts was $50-$100 depending on who’s fabbing it and how quickly it’s needed. Please note that this guide is only intended as a launching point for your simulation efforts and is by no means comprehensive. Hope you learned something!

• **This newsletter was sponsored by Onshape.

Hope this week gives you a real break from the grind—good food, good people, and a little time to recharge!

🏭What we have been up to!

4 years ago, we started Hardware is Hard locked in an Airbnb with COVID, purely as a fun, informative website to help other students at Berkeley. Over the years it has grown and grown, and we have loved the journey!!

While it has been a side project, we always thought that one day it might come in handy for a future endeavor.

Today, we’re calling in that favor!

Six months ago, we both left our W-2 jobs at big tech companies (Anduril and Apple) to start a manufacturing company in the defense space here in America! It has been a whirlwind and an absolute blast! We aren’t ready yet to fully share what we’ve been up to with the world, but in short time I’m sure a whole new portion of Hardware is Hard will be dedicated to telling the story—and all the lessons along the way—so that many of you can also start manufacturing companies here in America . . . stay tuned.

In the meantime, we have an ask.

We are delivering our first product and are looking to expand the business.

We are looking for personal introductions to the companies listed below. Trust us—we’re already doing all the cold outreach (it’s practically the same as getting an internship or job). However, if you personally know someone who works at one of these companies in a MechE, EE, or Supply Chain role (a friend, old coworker, neighbor, college floor-mate), we would be forever grateful if you could make an intro.

Please shoot us an email at hardwareishard@gmail.com, and we will set up a call to explain what we’re looking for and see if an intro makes sense!

Companies:

Saildrone - Vatn Systems - Teal Drones - Chaos industries - Jaia Robotics - Electric Sheep - Ulysses - Auterion - Quantum Systems - PDW - HAVOC - Picogrid

Thank you, and we’ll keep you posted on the endeavor!

OK, back to the regularly scheduled content!

⚙️ Fast Transparent Part Sourcing 💨

It can be super frustrating as an engineer to use parts marketplaces (-cough cough- Xometry) where you don’t get to talk with the actual shop making your parts. The “black box” situation can cause confusions, delays, and long term quality control problems (what if the shop they use for your parts changes between orders?).

Jiga solves this problem by helping you source high-quality, cost-competitive custom parts faster by partnering directly with vetted manufacturers.

And . . . it’s working. THEY JUST RAISED A $12M SERIES A

Try Jiga Today

👶 Meme Of The Week

🧠 Part Time MechE Internship (During The Semester)

Stell is a seed stage start up based in Venice Beach, CA. Stell makes requirements management software for hardware/aerospace companies - helping satellite, rocket, ship, and drone companies manage complex engineering projects.

Stell’s CEO and AI engineering team are looking for a part time engineering intern to help us develop a database of engineering standards data (like MIL-A-8625) in the Stell format, so that it can power our AI agent to develop templates and draft documents for hardware engineers.

Offering $25/hour. Work cleaning the documents is honestly boring, but the idea is that you are coming up with improvements to the product that make it faster/better. The impact to Stell’s customers is HUGE!

Looking for someone in mechanical/aerospace with an interest in AI and/or software engineering. Ideally looking for someone already local to Los Angeles that could come to the office. Ok to be remote during the holidays. US Persons only (ITAR).

Apply Here

🙋‍♂️ Interview Practice Question of the Week

Company: Tesla

Lights out and away we go! It’s Furnando Alonso’s first race with Aston Martin. As he heads into the first turn (radius 45 m), he slows down by 80 mph over 2 seconds and continues decelerating at that rate.

At some point during the turn, his car’s (mass = 800 kg) velocity has slowed to 60 mph. Have his tires (μ of racing tire on dry asphalt = 1.6) maintained traction or is he out of control?

Note: Please show work for and explain all steps. This is a 5-minute question.

✅ The Answer

Based on the length of the question, you can often tell if there’s a trick to it. For example:

• A very short question is likely to have multiple answers and require some level of deeper thought. It may be meant to make you think it is short, sweet, and simple. That’s usually deceptive (and even more deceptive when it actually is easy). Always check assumptions.

  • The majority of technical interview questions actually sit here, at only a sentence or two. The art is in answering it completely and can take 10-15 minutes if thorough with follow-ups in some cases.

• A mid-length question with some numbers thrown in there can definitely cause the pulse to rise. Here, we need to actually do the math. It’s not all that common, as most interviewers prefer to just leave variables undefined, but it does happen. Just take some deep breaths, start with what you know, and go from there.

• A lengthy and complicated interview question is very rare but does occasionally come up. If it’s too long to do in the time given, too complicated to be reasonably expectable or seems exorbitantly computationally heavy, you’re probably missing something. It is almost always a trick question. Check the assumptions, figure out what you missed the first time through, and swim through the red herrings.

This interview question is a mid-length question that is exceedingly reasonable to be asked on the spot and uses only basic physics knowledge (interview questions are usually > 20% easier than the ones you encountered in your classes to account for nerves).

Start with a free-body diagram (FBD).

Start by finding common units, I highly recommend metric.

60 mph = 26.8 m/s | 40 mph = 17.9 m/s

Next, calculate the rate of negative acceleration:

a = -17.9 m/s / 2 seconds = -8.95 m/s^2

The basic principles at play are friction and centripetal force. When the centripetal force overrides the frictional force between the tires and the ground, traction will be lost. This should be common sense, as friction is what is allowing Furnando to steer in the first place.

Equationally then, we can simply find the force of friction (Ff) and centripetal force (Fc) and see which is greater.

Fnormal = 800 kg * 9.8 m/s^2

Ff = Fnormal * μ = 800 kg * 9.8 m/s^2 * 1.6 = 12,544 N

Fc = (m*V^2)/r = 800 kg * (26.8 m/s)^2 / 45 m = 12,769 N

And so we find that Fc > Ff and Furnando Alonso’s start to the season suffers under Aston Martin (perhaps as predicted).

If you and/or your interviewer were/are on a Formula SAE team, they may ask some practical follow-up questions here:

• What factors did we not consider when calculating Furnando’s ability to make the corner?

• As Furnando begins to lose control (assuming this is a right hand turn), which way should he turn the steering wheel to avoid spinning out into the gravel?

• What are some ways that we could have allowed Furnando to maintain the same speed profile without losing control?

If there’s one piece of content we’ve written that you should know, it’s our technical interview guide. We recently revamped it with more info on materials and statics, check it out here!

👶 Meme Of The Week

⚙️ The A+ MechE Project 🗝

Having a good project to showcase in portfolios and discuss during interviews is essential for any mechanical engineer. Coming up with a strong project idea and successfully executing it can be quite challenging. All too often, mechanical engineering students either have ambitious project ideas that never get completed (or started for that matter) or choose something that turns out to be more of an “I followed an Instructable” endeavor (like building a classic motorized skateboard). In this guide, we aim to walk you through our thoughts on how to ideate and execute a successful project, accompanied by an example from a project we completed a couple of years ago.

The Idea

Generating a solid project idea might actually be the most challenging part of the entire process. You need to select something that you can realistically complete within a six-month timeframe (one semester), have the necessary tools to build, and produce results that are measurable and demonstrable in a portfolio and presentation.

Good ideas don’t necessarily have to be groundbreaking or entirely unique. We believe that good projects exhibit three key characteristics:

1. Exhibits Technical Hardware Design: This can encompass a wide range of things and doesn’t necessarily require reinventing the wheel. The challenging hardware aspect could involve packaging something relatively simple into a small, aesthetically pleasing form factor that is highly presentable. A prime example of this was a project undertaken by some of our classmates—an automatic chess-playing robot. In reality, getting an X-Y 3D printer-like mechanism to work should not be a challenge for junior or senior mechanical engineers, but packaging it into a normal-sized chessboard that is only a couple of inches tall presents a more difficult constraint.

2. Incorporates Electrical Components: We live in an era where mechanical engineers are expected to be multidisciplinary. It’s not sufficient just to design parts; you need to make them move. Additionally, when it comes to demonstrations, everyone loves something that can actuate! From a strictly mechanical standpoint, manual designs are more technically pure and some of the older generation live and die by this philosophy (looking at you unnamed Professor). However, when it comes to wow factor, it’s hard to beat the magic of automation, and it is guaranteed to make your project more interesting/appealing to someone reviewing it.

3. Uses Clever Code: As a mechanical engineer, you’ve probably dabbled in Matlab and Python at some point in time. Integrating these skills into your project is crucial (yes, you can find someone else to do the coding, but we recommend doing it yourself). Your electronic components won’t be able to do much without some coding, and a static robot is a sad one. That being said, we don’t expect you to produce perfectly streamlined, efficient code. Personally, we subscribe to the philosophy that computers are just rocks we struck with lightning and tricked into pretending to think. A hack is only a hack if people can tell it is. 95 times out of 100, no one is going to look at your code base but you, so comment your code however you want, use 10 if statements in a row, give your variables fun and meaningless names. No one will ever be the wiser [probably]. All we care about is that it functions to an acceptable level (depending on the project requirements, this may be rather more involved than we’ve painted. Either way, no need to go learn C in order to scrape the bottom of the barrel for an extra 50 milliseconds of performance unless necessary for basic functionality).

Bad Ideas:

• The Drink Machine: Getting a couple of solenoids to actuate is not very challenging, and this project is quite common. Without truly exceptional industrial design and a high degree of functionality, this project usually ends up looking like keg that sort of works.

• The Electrified Skateboard: There isn’t enough technically challenging meat to sink your teeth into; you simply purchase the motor(s), ESC, and battery and assemble them. We are not impressed. A good project needs ample opportunity for things to go wrong! Try throwing in a differential or making the deck out of composites and suddenly there’s a whole lot more to talk about.

• The Pancake Machine: A 2D printer with a batter dispenser is more than likely to end up a mess. If the controls and software are executed well, it could be interesting, but most often, the outcome is rather lackluster. Generally speaking, working with food can be tricky since it requires consistency of input and organic shapes aren’t precisely geometric, they’re….. organic.

• A 3D Printer: You can (and probably should) buy one for $200 on Amazon. Try for something more challenging! If you’re into making tools, a CNC would be a better choice as the design space is now force constrained making the challenge substantially more difficult. When you finish most good projects, you’ll have two sizable lists: things that went wrong the first time, and things that you want to change/improve. Why not make a 3D printer/CNC combo while you’re at it!

• A Quadcopter: Assuming you buy the parts off the shelf, on its own, it’s not a novel project- similar to the 3D printer. However, building an EVtol or using a quadcopter as a platform for another project can be worthwhile.

Good Ideas:

• The Aesthetic Chess Machine: This offers a great opportunity to create sleek mechanisms with “clever” code components that may be sourced from libraries, but the end result thinks on its own! Form factor and implementation are everything here, allowing you to execute on some beautiful and creative design choices!

• A Handheld CNC Router: This project necessitates a machine that can compensate for your movements (clever code), withstand the high cutting forces, and be in a user-friendly form factor.

• A Homemade Injection Molder: This project involves handling high forces, precise heating control, and producing a “finished product” that you can present and capture in photos. It also involves firsthand experience with tooling design and mold defects! Two things which any company doing plastic part design will go absolutely nuts to hear you talk about your experience with in an interview.

• An Automated Fire Extinguisher: Combining a cool actuation mechanism with computer vision components for a compelling demonstration.

• A Custom Carbon Fiber Bike Frame: While this project may not possess all the characteristics we mentioned, it is unique, involves intricate analysis, and demands challenging manufacturing processes, making it an exception.

Simply put, a good project is one where things can (and do) go wrong, there is a clear objective with freedom of implementation, and you learned something! A bad one is one with little to no creativity (it’s hard to design something new when the design space is so limited), is too easy (or too hard for that matter - infeasibly difficult project ideas sit on the backburner long after they should’ve been tossed), or is boring. It’s a personal project, so get personal! No matter what it is, make sure it’s something you’re interested in. If it’s not, the fight will be uphill from the beginning.

Taking Stock

Understanding your inputs is essential in constructing the framework to build a successful project. You need to choose an idea and scale that matches the resources you have at your disposal.

• What tools do you have access to? Can you get access to a 3D printer, laser cutter, waterjet and/or CNC? How much volume can you handle with these tools (how many iterations could you complete for your project)? If it’s a friend working in the machine shop that’s doing you a favor, making them crank out 6 iterations of a 3 op piece would be a tad overkill. Iterate faster using 3D printing where applicable, then upgrade for the final version once the geometry and setup has been validated. This will help save money, material and favors for another day.

• What is your budget? Selecting a project that requires large motors and substantial amounts of metal probably doesn’t make sense if your budget is less than $200. This is typically one of the more challenging decisions as most personal projects for young engineers tend to function on shoestring budgets. We would encourage you not to pinch pennies too hard on a good project. If it’s something you’re excited about, consider it an investment. Compared to the tens of thousands of dollars a year you probably spend (or have spent) in tuition, you’re likely to learn 100x more per dollar invested on a good project (and it’ll be more valuable for your career). Purely looking at the return on investment (ROI), spending a bit more on your project if you can afford it is a deal you can’t afford to pass up. Plus, after you’re done, many of the components will be able to be reused in your next project!

  • If you’re especially concerned about budget, there are often grants provided by your school for projects, estate sales, or junkyards you can pillage. Repurposing parts from broken or cheap secondhand electronics can save a shocking amount of money. My personal favorite item to deconstruct and ravage for parts are used microwaves. The number of microwaves available on the street, Craigslist or Nextdoor is near infinite. In them, you can find motors, large capacitors, LEDs, a magnetron, and a high power transformer (among other things). These items can be used on any number of projects. Another popular item are the motors used in car door windows! Crack one open at a junkyard and go crazy! Oftentimes, inspiration for a good idea will strike once you have parts in hand. Getting ideas at a flea market or junkyard is a great way to do that! Plus, nothing makes a bad day good like a little deconstructive engineering.

• What’s the timeline? A common mistake made by many students in project classes is waiting for the professor to provide the rubric and say “go” before starting on a project. This is a significant error because the first 3-4 weeks of the semester are usually the least busy and offer a fantastic head start on a project. Don’t let yourself make the excuse “Oh, I had midterms” when you’re two-thirds of the way through the semester with nothing to show for it.

• How many members does your team have? In reality, 1-2 team members often do the majority of the work in a team project, so having a team of 10 people is usually excessive. We’ll revisit this later when we discuss how to divide the project.

CAD Software

The choice of CAD software can make or break a project. If there’s one lesson we learned from team projects in college, it’s that Solidworks DOES NOT WORK. You won’t have PDM (Product Data Management - basic CAD filesharing software) running, and everyone will come to a meeting with CAD work done in different software (one person using Solidworks, another using Inventor, a third using Tinkercad or something random unfortunately), and after 20 minutes of downloading and uploading files to get everyone’s work into the same assembly, nothing will align.

The units will be off, and the parts won’t fit. It’s a DISASTER. That’s why we have one rule for projects: they must be done in Onshape.

If you want to join our project team, you have to use Onshape. There is no other option. Think of Onshape as the Google Docs of CAD. It’s free for educational purposes and allows everyone to work in the same document. It runs in a web browser, so people can use Macs, PCs, and even Chromebooks. It’s a game-changer for group projects.

The revision history works like Google Docs, allowing you to see the changes people make and when they make them. This helps with accountability. You’ll know how much time people are investing in the project and if they rush to complete it at the last minute. You can also collaborate on the same document simultaneously without overwriting each other’s work. This is crucial for CAD work and collaborative efforts. You can even leave comments and store important PDFs in the CAD workspace. As you’ll see throughout the rest of this guide, all of our work was done in Onshape. The best part? It’s free for hobbyists and students! Get an educational Onshape account here.

Start CADing Today!

Execution

Once you’ve selected a project and assembled a team, it’s crucial to create a roadmap with deadlines and realistic expectations. Everyone should mark their midterm dates on the calendar to get an idea of when you’ll have time to execute. With a roadmap in place, it’s essential to shift your focus to risk management.

• What are the unknowns? Keep in mind that there’s a big difference between not knowing how to do something, and not knowing if something can be done. The former presents only medium risk, where the latter must be addressed immediately as it could have downstream implications for the project as a whole. Keep in mind that when there are things you don’t know, there’s probably also at least that many things you don’t know you don’t know.

• What aspects should you validate early to prevent going too far down an unproductive path? Are there a few different architectures you’re considering? Try doing a small sprint to evaluate the predicted merits of each. There’s not time (nor willpower) to do a full parallel path, but checking to compare a couple different architecture proof of concepts can let you quickly choose if not the best, then at least a viable path forward.

• Isolate. Isolate projects into compartments with concrete deliverables. What parts of the project rely on each other? Sometimes it’s difficult to test what will or won’t work without another functional component. How do you test a touchscreen without programming the GUI? Or controls without a robot to control? These blockers can often be circumvented by creating scaled or representative models (or digital models if you’re so inclined), but keep in mind that it costs extra (potentially unnecessary) work up front, so keep project flow linear if possible.

Usually, completing a Minimum Viable Product (MVP) before midterms is crucial for a successful project. What’s an MVP? It’s the most basic form of your complete idea. It can be assembled with hot glue, 3D printing, and even cardboard. It doesn’t need to look attractive; it just needs to demonstrate that the idea can work and provide insights that will help you refine the final product. Your MVP should serve as an integration testbed that allows for early testing of programming and electronics, rather than waiting until two days before the final deadline (which happens far too often).

One of the most significant challenges in team projects is member participation. Typically, 1 or 2 people do most of the work while others do the bare minimum to secure a grade. This is unfair but often partly the fault of the 1-2 active contributors.

Firstly, no team project should have more than 5 members. It’s practically impossible for 6 or more people to work effectively on a single class project and all make meaningful contributions. For groups in the 3-5 range, the key is to divide the project into verticals that each team member can take ownership of. Each member must feel directly responsible for a deliverable. A common approach is to have 2-3 individuals focus on mechanical subsystems, one on electrical aspects, and one on software. The electrical person should make all the moving parts work independently of the mechanical system, and the software person should collaborate with this electrical testbed to mitigate code-related risks. One hotbed issue with splitting up responsibility for deliverables is trust. Making people responsible for individually critical verticals means introducing multiple points of failure. Now, someone not pulling their weight isn’t just 20% more work for everyone else, it’s catastrophic failure. I would like to say that people will rise to the occasion, and have often found that to be true, but I will also say that there is always the distinct possibility that they will not. That is why work parties and status updates are critical. Don’t let someone get away with generic statements like “looking into motors.” Make them show you the different options and talk about the benefits of each. Dive into their codebase. What functionality has been coded so far? Has it been tested? Nothing is worse than testing code for the first time with only 2 days left until the project deadline. This should be avoided at all costs.

To rectify many of the issues and risks presented above, we also highly recommend “work parties”—3-5 hour blocks where all team members are fully present and engaged with project-related work to ensure progress and status updates on all fronts. Members ideally should not be working on homework for another class, on Instagram, or studying for midterms. Each work party should start with a 3-5 minute update from each group member in which they present slides on their current work status, updates since the last meeting and any potential risks/blockers they have been dealing with. Group input may be able to resolve those blockers, and shame of presenting nothing to the group will force them to do at least some work. It’s impossible to hide in the weeds when you have to present on work you alone did, and many people tend to find this motivational to keep the ball rolling on their individual deliverables. This also mimics an industry environment in which you give technical updates in 1:1’s to your manager or to the broader team. One vital piece of being able to present on those deliverables is documentation!

Documentation

One of the most critical aspects (possibly the most crucial) of a project is proper documentation and an effective final presentation. If you didn’t document it, it didn’t happen. This is especially important for those who want to include the project in their design portfolio. Throughout the process, take screenshots of CAD work, photograph both successful and unsuccessful prototypes, and record your thoughts as the semester progresses. We recommend creating a mid-semester presentation document with all the updates on progress up to that point. This ensures that you don’t forget about all the hard work done early on. At the end of the project, a well-structured presentation is essential. It’s all about storytelling. What challenges did you tackle? What unique design innovations did you develop? Why is the project a resounding success? Often, the grade you receive or the impression you make on an employer depends more on the presentation and framing than the actual project itself—presentation is everything!

We actually wrote a bunch more about an example project from our senior capstone. Read more about it here!

MechE Projects 101

We put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

🚨 Optica Industries is hiring!

They just raised an $8.3M seed round and are looking for a founding Mechanical Engineer to help them build a metal processing facility capable of rapid and accurate sortation of base metals for recycling. The job involves industrial automation, sensor integration, design, PLCs + microcontrollers, welding, pneumatics and structural simulations.

If you are passionate about America and its material base, and love unsexy industries, this is the job for you (located in SF). Reach out to sam@optica.industries if interested!

🦸 Guest Writer: Sajiv!

This week’s newsletter is a special edition written by Sajiv Shah about takeaways on manufacturing in China after his recent actuator vendor visit! Sajiv is a 3rd year Mechanical Engineering student at the University of Michigan and has racked up some impressive internships already. In past internships he has worked on humanoid robots at Figure AI, rockets at Relativity Space, and electric propulsors at Whisper Aero.

In the last decade venture capital investment into deep tech companies has skyrocketed. It finally feels like the playbook in the Founders Fund manifesto is starting to see the light of day, as specific venture firms are beginning to pour large amounts of capital into moonshot ideas. “Re-industrialize America” is a common term I hear at startup fairs, and it feels like it is one of the best times to be a Mechanical Engineer who wants to work at a deep tech startup. Undoubtedly, one of the driving forces behind this push towards more hardware development in the U.S. is the political climate around relations with China.

A few weeks ago, I had the unique opportunity to visit China with a group of robotics enthusiasts and walk the factory floor of various component manufacturers. It was simultaneously fascinating and anticlimactic.

First Impressions

I got my first glance of the scale of Chinese cities during my descent into Shanghai. Everywhere I looked there were large skyscrapers that seemed to continue for miles and miles, and further from the airport they were densely packed next to farmland. I’ve had the pleasure of marveling at skyscrapers in Manhattan, San Francisco, and Los Angeles out of plane windows, but Shanghai felt different.

It seemed like everywhere we went there was construction. At any moment looking out of a car window I would see skyscrapers being built or sections of high speed railway being assembled. China built their first high speed rail in the late 2000’s, decades after other countries, and now has an order of magnitude more rail length than the next few countries combined. They have so many skyscrapers that vacancy is becoming an issue. Different regions are competing to become the next major hub and will continue to industrialize to no end.

Shanghai & Shenzhen

Our six days in China were split between Shanghai and Shenzhen, which showed two very distinct identities of Chinese manufacturing within the robotics space. Our primary purpose of visiting factories was to understand who the major players are in manufacturing robot actuators and what direction these companies envision the robotics industry heading toward.

The three major companies we visited in the Shanghai region were spun out of larger incumbent companies which produced motors or gearboxes. With the growth of industrial robots in the 2000’s and 2010’s, the Chinese government made several investments into traditional automation, leading to the creation of multiple companies that produce planetary and RV reducers. Within just a few years, China surpassed Japan—where these reducers were originally designed—in market share.

Robot actuators primarily consist of the same gearbox components but at smaller scales. Incumbent companies view humanoid robots as an early market with enormous growth potential. It’s clear they are trying to get ahead of the curve by applying their expertise in gearbox and motor manufacturing to humanoid-related R&D efforts.

On the manufacturing floor, you’d find exactly what you would expect: large rows of specialized machines cutting gears, turning parts, and grinding. Most were human operated, but a few had industrial robots picking and placing items—something all the manufacturers want to invest more in to reduce labor cost.

The culture at these companies is starkly different from silicon valley startups. Most engineering and business leads we spoke to were older, and had worked their way up from the manufacturing floor. At one company, every engineer was required to spend time on the manufacturing floor when they joined the company. Notably, none of these companies believed that they would drive the market with their new products. Instead, they were focused on building out capabilities such that they could quickly respond to changes in demand once the robotics market grows.

Shenzhen culturally felt like the San Francisco of China. It is a very young city, which just over 40 years ago was not much more than a fishing town. Its growth began with its declaration as a Special Economic Zone, combined with foreign investment due to its strategic location at Hong Kong’s border.

Its surge, however, is largely attributed to the growth of its electronics manufacturing base. This was fueled by consumer technology companies like Apple, Sony, and HP outsourcing manufacturing to Shenzhen-based companies like Foxconn in the early 2000s.

The companies we visited in Shenzhen were not too un-like silicon valley startups. They were run by young founders, venture backed, and operated out of incubators and accelerators. They were chasing new markets and competing with incumbents with more innovative methods.

The ability to scale a company in Shenzhen is incredible, with close access to a labor force and components for prototyping (the electronics markets were remarkable, I felt like a kid in a candy shop).

Ecosystem

In the states we have a strong notion that we used to be the leaders of industrialization and manufacturing, which primarily points to WW2 and the cold war era. How did China emerge? Why did we fall behind?

Arthur Kroeber on the Dwarkesh Podcast gave a great explanation of this. In the 1990s, China identified automotive prowess as a key to economic development and technological growth. The government spearheaded joint ventures between German automakers and Chinese manufacturers, spurring rapid growth—but failing to produce truly global carmakers.

In the late 2000s, ahead of their boom, the Chinese government identified EV’s as the next frontier technology and decided to heavily invest in the industry and support companies like BYD. Over the next decade several companies had solidified supply chains and technology in battery manufacturing, but it wasn’t until the growth of Tesla’s vehicles in China that Chinese suppliers improved consumer design and started to dominate the market.

Today, these automotive companies build products across many different disciplines and are highly innovative. Xiaomi makes everything from phones to cars, and plans to produce humanoid robots. As we saw in Shanghai, incumbent companies are eager to follow government funding and launch into new product sectors to be competitive. In terms of the premise that automotive companies help growth into new business ventures, China made a very good bet.

The U.S. followed a similar pattern during our industrial peak. In the 1940’s, Ford was rolling B-24 bomber aircraft off their manufacturing floors. Chrysler was developing rocket engines, and General Motors was producing ammunition, aircraft engines, and tanks.

When wartime and government funding called for production, American automotive companies stepped up the challenge. Today, most of these companies fell behind the EV race. Tesla exists as one of the only prominent automotive companies in the U.S. that is expanding into robotics, chip manufacturing, and AI.

Takeaways

I came into this trip with the assumption that China succeeds because of lower labor costs, hosts an abundance of valuable resources, has special capabilities to manufacture scale, yet still lacks the engineering dominance we have in the U.S. While it’s true Chinese labor is cheap and resources are plentiful, modern engineering efforts are narrowing the gap. On the floor, there are no special machines or processes that a manufacturing engineer can’t imagine. And while it is true that most progress in AI/Robotics/Self-driving technology is coming from the United States, we met plenty of highly talented young engineers in Shenzhen–like the founder of the Wuji robot hand, arguably the best robot hand available for purchase today.

How will things change with AI/robotics?

A common fear Chinese component manufacturers expressed was that if the U.S. develops highly capable humanoid robots, demand for their capabilities will diminish. This fear initially seemed irrational, given that most humanoid companies like 1x and Figure AI are focused more on general-purpose home robots than factories. However, with the growth of adaptable and intelligent non-humanoid form robots enabled by AI, I imagine many medium volume manufacturing/assembly processes will be kept domestic in the U.S. as labor costs can be cut while process control and turnaround times can be improved.

What is certain is that China is heavily invested in robotics. Every company we talked to was very excited at the prospect of working with American humanoid companies. The margins and demand from overseas companies dominates what they foresee from local developers like Unitree.

Visiting China gave me an opportunity to step back and think about how differently the world’s two largest economies developed over the last few decades. The growth of Chinese manufacturing is proof that top-level investment into an ecosystem and competition will drive innovation and growth at unparalleled rates. If you get the chance to visit, I highly recommend it!

We have put a dozen guides for mechanical engineering students and early professionals on our website

• 50 Hardware Startups who have raised less than $50 million (perfect internship targets)

• How to handle The Behavioral Interview

• What it takes to be The 10X Intern

🙋‍♂️ Interview Practice Question of the Week

Company: Cover

Which is stronger, the screw or the nail?

✅ The Answer

This question is incredibly vague. That’s intentional. To answer this question, you need to first ask questions.

Primarily, what direction and type of load is it? How do the nail and screw size to each other? As always, when you ask this question, the most likely answer is: “Explain for each.”

Firstly, let’s handle the loading. There are two main types of loading we’re going to consider: tensile and shear. For tensile loading, the usual calculation is:

Stress = Force / Area

So a higher area = a lower stress. From a purely theoretical perspective, this might be correct. Most of the time though, nails and screws are used in wood. As such, let’s assume we are using this screw/nail to connect two blocks of wood. Now, we must check our assumptions. In most tensile loading questions, there is an underlying assumption that both ends are rigidly held onto (i.e. there is no slippage). This is true for one end of the nail (its head rigidly makes contact with the block, preventing any further ingress. The other end of the nail though, lacks the same feature, meaning it can be pulled back out of the block if friction is overcome.

In short, if Ftensile > Ffriction of block 2, then block one and block two will be pulled apart, regardless of the diameter of the nail/screw.

A screw on the other hand, uses threads. So long as the tension-generated radial force due to helix angle of the screw threads does not exceed Ffriction, the screw will not rotate (a frictionless screw would rotate back out and be equally as useless as a frictionless nail). As such, any tensile forces on the two blocks are not limited by friction, but rather by the strength of the wood, strength of the screw, and thread type of the screw. Deeper threads will increase engagement with the wood, meaning that the wood has a higher shear area to resist tensile loading. Stronger wood will have the same effect.

An additional failure mode to consider, is that the screw could snap between the blocks, allowing them to separate. Some semi-common knowledge that could be useful to bring up here is the difference between cast and forged components. Forged components are significantly less brittle and tougher than their cast counterparts due to adjusted grain structure and lower porosity. For this reason, forged fasteners are often used in load critical applications.

An article well worth the read is NASA’s Investigation of SpaceX’s CRS-7 Launch. The problem boils down to using cast rod ends instead of forged ones, which broke under vibrational loading (at low temperatures) 139 seconds into the flight causing the COPV (think big carbon fiber balloon held under water) to shoot up into the tank and rapidly deconstruct the rocket in less than a second. Reading the report, it appears that the manufacturer suggested a factor of safety > 4 for using the cast rod ends. This should give you a good idea about the structural inconsistency of some cast components.

🚀 Hestus - CAD Copilot

Reclaim your design time. Hestus is the AI copilot for Autodesk Fusion that eliminates the most tedious parts of your workflow. Instead of manually applying dozens of obvious sketch constraints or dimensions, Hestus proactively suggests them for you. Spend less time on digital housekeeping and more time solving the complex problems you love.

Watch the demo and try Sketch Helper for free at www.hestus.co.

Share with us how you’re using it in your designs — the first 10 entries snag a $50 Amazon Gift Card!

For shear loading, what we care the most about is the cross sectional area of the nail / screw. It is very important to note that the cross sectional area of the screw we must consider is the minor diameter. This is the narrowest diameter of the screw, and where failure will occur. It could also be argued that a screw will better resist shear for this particular application of binding two blocks together, as the friction between the two blocks will ideally play a role (this is actually how bolts work, and it’s very interesting — hopefully we will release more on that later). That would however be slightly dependent on if one of the blocks was predrilled at greater than the major diameter. More generally though, the problem of shear loading performance can be answered by checking the minor diameter of the screw vs. the minor diameter of the nail and seeing which one is larger.

Note: the designing of bolts is an underappreciated art. There are all sorts of standards dictating every part of a screw for different purposes. One of the most striking ones (and my personal favorite) that we rarely think about, is the small flat portion on the inside of the threads that helps prevent crack propagation.

Torsional loading doesn’t really need to be discussed in this use case, as the system is under constrained without a second point of attachment (which would almost always be added). By adding a second point of attachment, the torsion turns into two cases of shear!

The summary regarding which is stronger: a nail or a screw? As with pretty much anything, it depends! Primarily on the use case. Hopefully this response has helped you with the tools to analyze any use case you are given.

Now when it comes to practical knowledge, there are some general rules of thumb. For example, nails tend to be less expensive than screws so, for things that require buckets and buckets of them like roofing or framing, nails are probably the better option. They are also quicker to install so good for keeping UPH (units per hour) up. Screws are best for cabinetry (unless you got a brad nailer) and flooring because both things see lots of loading and you don’t want squeaky floors or loose cabinets.

Also, DO NOT FORGET to pre-drill your holes. Otherwise, you may crack your wood. It’s a tedious step but nothing worse than finding a crack in your wood just when you thought you were done.

Thanks for reading! Have something interesting (and hardware related) YOU want to write about? Drop us a line at hardwareishard@gmail.com!

This week’s newsletter is a special edition written by Madhav Bhat with input compiled from several other engineers all with recent experiences in the internships —> full-time pipeline! Madhav graduated from the University of Michigan with a BSE and MSE May of 2024. Throughout his recruiting journey he interviewed for mechanical engineering internships and full-time positions with companies from large Fortune 500 corporations (think Ford, Michelin, Apple) to growth-stage companies (think Tesla, SpaceX, Anduril) to startups ranging from 3-1000 people in size working on all sorts of projects.

Check out our last post for part 1 (Design positions in different industries).

Manufacturing Roles:

In manufacturing, your role is entirely dependent on the product and the manufacturing volume so that’s why I’ve broken this section down according to volume.

🤠 Meme of the Week

Established Companies:

When you’re entering a well-established company, your role is often more structured than at a startup or small shop. The systems, processes, and workflows are already in place, and your job is usually to operate within them, improve them, or specialize in a narrow slice of the larger machine.

One key difference is that you’ll often have clear ownership over a particular function, subsystem, or set of processes. Instead of wearing several hats, you’d be responsible for something like maintaining a family of processes related to components such as a glue dispense, attachment biasing and related machinery specs/updates. Oftentimes, goal is to increase yield or lower units per hour (UPH). The upside is that you’ll gain deep expertise in that area and leverage with vendors is high, but the tradeoff is less variety compared to early-stage or small companies.

In these environments, you’ll spend a lot of time working with established procedures and standards: ISO, AS9100, Six Sigma, DFMEA/PFMEA, and other structured frameworks are the norm. Learning these systems and demonstrating you can operate within them effectively is often more important than raw creativity or output. That said, incremental improvements—reducing cycle time, eliminating a source of rework, increasing yield—can have a huge business impact because the scale is so large.

Another defining feature is the cross-functional aspect. At scale, engineering rarely operates in isolation. You’ll work closely with not only manufacturing compatriots, but supply chain, PD, purchasing, and quality to ensure design changes are cost-effective, manufacturable, and reliable. Being able to navigate these handoffs smoothly is just as valuable (if not more so) as your technical skills.

You’ll also want to develop a strong sense of documentation and traceability. In established companies, “If it isn’t written down, it didn’t happen” is basically a rule. Specs, reports, test plans, ECOs, CAD revisions are your bread and butter. Learn how to make your work clear, defensible, and easy for others to pick up. As with literally everything else in engineering (and life?), communication of information is pretty much just as important as the information itself.

Finally, keep in mind that career growth in these environments is often about demonstrating reliability and building credibility over time — or at least that’s what your HR representative will say. IRL, it means “years in role,” which is a really convenient excuse to be passed up for promotion regardless of skill level. That said, it also means that if you stay long enough, you will get promoted up to a certain point. Past the usual plateau point though it requires a little more pizazz to get promoted just as with anywhere else.

Medium to High Volume:

A Design for Manufacturing course should have a good basis for the different manufacturing methods you might encounter, like injection molding, stamping, extrusion, die casting, etc. A good thing to know would be what manufacturing method to use for certain types of parts, why, and the benefits and trade-offs between manufacturing methods.

Your job will most likely focus on reducing errors in the manufacturing process. When identifying a problem, focus on both the criticality and frequency of the issue. Brush up or learn some statistics as statistical process control is huge in the field. This topic came up quite a bit in interviews and it makes sense: when manufacturing at scale you are dealing with 100s to 1000s of items on the line and a variety of factors/variables. Using statistics to identify how these variables affect your output is key to optimizing performance. Learn how to design an experiment and conduct effective root cause and failure mode effects analyses (FMEA). These all boil down to how good are you at identifying the different ways things can fail and go wrong, and how can you prevent that from happening?

Also, don’t think you won’t have to be hands-on. You might have to design experiments and create test fixtures to carry them out. You might have to spend time on the manufacturing floor carrying out tests and identifying sources of variation.

One thing you’ll probably want to emphasize is communication skills. If you’re on the line you’ll be working a lot with the production team so try and win them over so that when you stop production to run an experiment they don’t get too upset… And be sure to clearly explain why the experiment you’re running benefits the team/company and has to be done. It’s also good to talk to them and identify what the highest priority issues are for them that are slowing production and then see if those align with the manufacturing team, sometimes there can be a disconnect.

Low Volume:

Think SpaceX Starship, Arc Boat, making a prototype at a startup, etc. You’re proving a concept or whatever you’re making is a huge engineering endeavor.

Low-volume manufacturing roles can look a lot different. At this stage, you will most likely be working on bringing concepts to reality. What this means is that you have to be very hands-on. Get experience building things with your own hands. Learn how to make things using a mill, lathe, CNC, waterjet, laser cutter, 3D printer, and whatever else you have in your machine shop. Working in small manufacturing environments at your university will be helpful so once again join project teams and get involved in hands-on projects. The startup advice that I gave in the design section applies here.

You’re probably going to be working in a fast-paced environment so minimal outsourcing. There isn’t really a trick to it other than just getting experience. Also it’s a good idea to have a portfolio of the things that you’ve built, how you built them, and why you built them the way you did. Another helpful idea is to brush up on various manufacturing concepts/processes across a variety of materials (plastics, composites, metals) as well as the adjacent processes like welding, soldering, heat treating, etc. and how that’s done.

🚀 PDX is The Hands-On Bootcamp for MechEs Who Build Physical Products:

• ⏥ Tolerance stacks & GD&T pro tips

• 🦾 Design for manufacturability (DFM) & assembly (DFA)

• 📊 FEA, failure analysis & root cause techniques

• 🧑‍💻 Advanced CAD surface modeling

• 🩹 Adhesives, coatings, and specialty finishes

• ✅ Building better supplier relationships & RFQs

• 🤖 Programming robots & PLCs for automation

• 📍 Oct 21–22 in Phoenix, AZ

Listed above are the main skills that any successful PD engineer must have. Whether you’re looking to level up your game or stay fresh with the latest software tools in the space, we highly recommend checking out PDX.

👉 Lock in and register here:

https://pdexpo.engineer

Use the discount code “HHR50” for $50 off!

Nontraditional Roles:

Consulting:

Mechanical engineering at its core is solving complex issues. I guess most of engineering is. But consulting firms value engineers because they are deeply technical and not afraid to get into the weeds and root cause issues. Consulting allows you to do that for not just engineering issues but a wide range of problems across all sorts of companies. To get some more insight on this I talked to Jonathan Jasica who is an associate at LEK consulting.

Why Engineering Students Have an Advantage

Recruiting for consulting is competitive. Unlike engineering roles, there are no technical barriers to entry. You will be competing against business majors, econ majors, CS majors, dance majors, etc. alike, but being an engineer puts you at a competitive advantage from the start. Consulting firms are looking for driven problem solvers that can attack broad business problems. So, problem solver, check. As an engineer, you have already gotten exposure to complex problems and have experience using math and technical skills to solve them. You will likely excel in this part of the job/interview. The part that most engineers find more difficult is understanding and driving to an answer in the broad business space. However, to let you in on a little secret, 90% of the time it boils down to maximizing profits (BTW profit=revenue - costs). It may sound different ways “they want to expand into X market” “they are losing $X a year” “they are seeing growth slow”, but at the end of the day it is often about how can the business get more dollars flowing in the door and mitigating dollars flowing out.

Building Business Knowledge

You will want to have a general understanding of the business world and trends, this will help you come up with ideas and demonstrate your knowledge during the interview process. If you are totally out of the loop of the business world–like I was–I suggest subscribing to Morning Brew’s newsletter or listening to daily business podcasts in the morning (Brew Daily, The Daily, The Best One Yet).

Recruiting Season

As with most jobs, getting the interview is the hardest part. Consulting jobs have high attrition and firms recruit full classes (depending on the firm, anywhere from a few dozen to several hundred new associates each year). Because of this high turnover, most firms have dedicated recruitment processes, internship programs, and university staff. Draft a strong resume. Use AI and friends to review and refine it when recruiting season begins–late summer for most firms. Highlight your gpa, ACT/SAT test scores, past internships/projects. Turn in a cover letter, but do not spend too much time on it (aka use AI), as likely no one will read it.

Network Strategy

Many people will tell you you need to network. I believe that is true, to an extent. Any firm you are serious about, try to talk to 1-3 people at it. This will help you understand the firm and give you some strong talking points during interviews (e.g., the ‘why us’ interview question). Anything beyond that, you are wasting your time and the conversations begin to be a chore, rather than anything valuable. Find them through your network, at recruiting events, or via LinkedIn.

Case Interview Preparation

Now it’s time to focus on interviews – the most critical part of your consulting journey. Like riding a bike or mastering calculus, interviewing is a learnable skill that improves with practice. Consulting interviews are particularly challenging, but mastering them will give you interview confidence that lasts your entire career. Start preparing well before the first interview request lands in your inbox.

The cornerstone of consulting recruitment is the case interview: a role-playing scenario where you solve a hypothetical business problem. This exercise evaluates your structured thinking, quantitative abilities, business knowledge, and presentation skills all at once. While it seems intimidating, approach it as a collaborative conversation rather than an interrogation.

To prepare, you’ll need practice partners and resources – but don’t get overwhelmed by the sea of guides, courses, and books promising success. Remember the consulting principle of “80/20” (the Pareto principle): roughly 80% of results come from 20% of your efforts. This means you don’t need hundreds of practice cases or expensive courses. Instead, find a few dedicated practice partners and use free online case books to develop your skills until you feel confident. Some useful resources here are casebooks from top MBA programs, like the Darden Casebook as well as other online case studies and university consulting clubs who may have a repository of this.

Interview Structure

Interview structure typically follows a first round consisting of 1 behavioral interview, with standard job interview type questions (use STAR technique), and 1 case interview. The first-round case interviews typically last 30 minutes and are given by a younger associate / consultant. They will be following a rubric that assesses your communication, math, analytical, business skills, as well as your confidence, throughout the case. The final round interview will look similar, just with high-level managers / partners giving the interview. They are more likely to go off script and ask you adhoc business questions. Don’t let that throw you off.

Consulting Career Path

So, you might be thinking, ‘gee, this sounds like a lot of work’. And, it is. And trust me, the job does not get any easier once you start. Consulting jobs often come with long hours, demanding project timelines, and monday-friday travel. So why do so many people want to do it? This is unique to everyone, but some common refrains are as follows:

• Salary: Consulting jobs pay well, often offering $100K+ straight out of college with yearly bonuses ranging from ~10-20% of salary

• Prestige: Who doesn’t love blasting their LinkedIn network with a sexy logo. Consulting firms are notoriously competitive and known for developing top business talent. Having that brand on your resume can open many doors

• Change: There is nothing worse than having a shitty boss. Consulting is not immune to that by any means. However, typical consulting projects last just a few months at most firms. This means every few months you will be working with a new team, in a new industry, with a new client. This gives you exposure to many different industries and projects and companies in a short amount of time. Constant change and constant new learning opportunities

• Career optionality: Perhaps the most common reason is career optionality. Consulting firms are known for developing business-savvy, analytical thinkers. Many people exit after 2-3 years into a plethora of fields and jobs. People go to F100 companies, startups, B-School. And enter job fields like corporate strategy, BizOps, product management, and private equity. And of course some people choose to stay and climb the ladder. Salary can be quite lucrative, should you choose to stay and get promoted to manager, principal, or partner.

It can be difficult to know what you want to do for the rest of your life as a naive 20-year old at college. Choosing a career path can be daunting. Consulting is a way of exploring different fields, job titles, and companies. You are an engineering major, you will be successful no matter what you pursue. It is not worth the effort just for salary and certainly not for ‘prestige’.

One other thing to note here, is that consultants often face some stigma within engineering. There is a general sense that you have to work on something for a couple years to reap what you sow. That is to say, jumping around to different projects and offering solutions sounds great in theory, but you don’t learn the same way that you would from offering a solution, having it fail miserably, and iterating until you find something that works. Especially true among seasoned engineers, the belief tends to be that you need to stick around long enough to be responsible for your own failures in order to grow to some extent. Also note that just because some people have some thoughts about consulting doesn’t mean you can’t convince them differently with your startling clarity of thought and communication.

Trust the Process

I was a mechanical engineering student at Michigan and around my sophomore year summer, I decided to try my hand at consulting. Hundreds of hours of mock case interviews and dozens of networking calls later, I have had the experience interviewing at many major consulting firms including BCG, Bain, EY-Parthenon, and I am now 6 months into my job as an associate at L.E.K. Consulting. Most applications ended with a simple rejection email, few led to interviews, and even fewer led to offer calls. The point being, this is a numbers game. No one rejection is any indication of your intelligence or performance. The fact of the matter is, there is a disproportionate number of open positions to overall demand. However, if you put the time and effort in, you will be successful.

Conclusion

Thank you for reading through my thoughts on recruiting and if they were helpful please share with others! The way various companies recruit is constantly evolving and although I’m not sure how relevant the specifics of this information will be in a couple years, I think the high-level principles in each section should still hold. Although I’m working more on the software side right now, mechanical engineering is one of the hottest fields to be in and it’s truly exciting to see all the new and cool technology being brought into the physical world. From portable nuclear reactors, to hypersonic planes, to humanoid robots, it’s an exciting time to be in hardware! Feel free to dm me on LinkedIn if you have questions/thoughts :)

This week’s newsletter is a special edition written by Madhav Bhat with input compiled from several other engineers all with recent experiences in the internships —> full-time pipeline! We love getting fresh perspectives and wanted to share Part 1 of this comprehensive How 2 Recruit that Madhav put together! Madhav graduated from the University of Michigan with his BSE and MSE May of 2024. Throughout his recruiting journey he interviewed for mechanical engineering internships and full-time positions with companies from large Fortune 500 corporations (think Ford, Michelin, Apple) to growth-stage companies (think Tesla, SpaceX, Anduril) to startups ranging from 3-1000 people in size working on all sorts of projects.

Introduction

During my years at Michigan, I realized that mechanical engineers don’t know enough about the recruitment process and how to recruit well. A lot of recruitment advice is outdated and we are entering a new age of engineering recruiting that universities don’t always tailor their advice for. The goal of this document is to fill in a lot of knowledge gaps and answer questions about recruiting for internships and full-time jobs at any company a mechanical engineer would want to join. I want this document to be the first thing a student references when starting their recruitment journey. Even though I did well at getting interviews, that didn’t necessarily translate into landing offers, whether it was “right place wrong time” or just bombing the interview. However, I did find a few tips with a high impact on odds of success.

I believe the “getting interviews” advice most appropriately correlates with the company growth stage which is how I’ve broken down the mechanical design section, so it is included there. Manufacturing (to be covered in the future!) is most appropriately broken down by production volume rather than growth stage so I’ve broken it down that way (startups tend to be lower volume, whereas growth and large companies vary). And then I’ve included some other semi-MechE roles at the end that students might be interested in.

This is general advice stemming from mine and some other people’s experiences so please take it with a grain of salt as it is anecdotal. I am not at all saying every company that falls into a certain bucket will recruit the same way; this is just general semi-specific advice :)

🤠 Meme of the Week

Mechanical Design Roles:

Like stated in the intro this section is broken down by company stage. Established companies include legacy mechanical engineering companies like big 3 automakers, mechanical component suppliers and more. Growth stage companies are slightly newer but have experienced rapid growth that has allowed them to become industry leaders and have different interview processes than legacy automakers. Examples include Tesla, SpaceX, Lucid Motors, and FAANG. Finally, startups are small hardware companies anywhere from Seed to Series E (and sometimes beyond lol).

Established Companies (Insights from Jack Decker former Mechanical Design Intern at KLA):

Recruiting at established companies can vary quite a lot, but I see in-person career fairs as the best shot at securing interviews. Recruiters on LinkedIn tend to be less responsive at these large companies; however, in-person conversations tend to turn into interview opportunities.

In Jack’s case, he met a recruiter at the career fair and attended an information session to learn more about the company and role. Following that, he had an HR screen and then a technical interview. I had a similar experience when recruiting for large automakers where a chat at the career fair led to the interview. The key factor here is passion. Do detailed research on each company before you go. Don’t just talk to a recruiter there, speak to the engineers and learn both about their day-to-day and cool projects. For example, Toyota’s engineering department prioritizes reliability (having started 6 sigma and all that), Lockheed might prioritize regulatory compliance and safety, Dyson might prioritize user experience, etc. Demonstrated interest can take you farther than you think here. Jack showed even more demonstrated interest by taking a PTC Creo course on Udemy when it was mentioned in the job description.

In both his and my experience, familiarity with the STAR interview format (situation, task, action, response) was crucial and expected. This is also just a good framework for all interviews. When discussing previous projects, be articulate, concise, and illustrative. Show why what you did was cool and how other people benefited from your project. Sometimes they will ask technical questions about things like root cause analysis or previous projects (What would you change if you went back? Most technically complex feature? Your design process and tools used?), but by and large these interviews tend not to get too technical. Expect project management and behavioral questions on topics like team and time management and prioritization.

For internships for most companies at this stage, expect about 2 rounds of interviews after the HR screen.

Growth Stage Companies (Insights from an Engineer at Joby Aviation):

A lot of times roles at these companies don’t involve just living in CAD, it’s more of a full service role. It can be multidisciplinary in nature involving design, testing, manufacturing, etc. An emphasis is placed on component ownership from inception through production, so whether you want to or not, you’d get to own all aspects of design from concept development to prototyping to test to production. How do you prepare for this?

The networking tips for recruiting are also the same (LinkedIn cold messaging and IRL networking). Many times, these companies will also show up at career fairs since they are hiring so much, but keep in mind that they will be the most popular booths and you will have to really stand out to convert into an interview. Be prepared to answer technical questions about past projects on the spot. They’ll also be at project team competitions like Baja and Formula SAE which are awesome recruiting events as well (but remember some of the most passionate engineers are at those events…so really think about how you’re going to wow these companies). I’ve gotten wrecked at a couple of these in-person events…

Also, look at the types of products made at these companies, what they prioritize in design, and how you can talk about your projects in a way that’s appealing to them. For example, in consumer electronics, a huge priority is aesthetics and mass manufacturability. In aerospace, your materials knowledge will come in handy. Etc.

In the Tesla application, they ask for “Evidence of Excellence” and that mentality translates across these companies. In my interviews with companies at this stage, I’ve been asked some of the most difficult technical and nontechnical questions which have made me feel embarrassed or question my engineering skill. I hope you don’t feel that way when interviewing, which is why I urge you to not only know the details of your projects but why you made the engineering decisions you did. This is where being able to talk about your ideas in a clear and digestible manner becomes extremely important. Many times the interview process in these types of companies include a technical presentation/deep dive that encapsulates this and also evaluates whether you’ll be a good coworker (can you communicate ideas well? Do they make sense? And of course, do you make intelligent technical decisions?). It’s not just going to be engineers that ask you these tough questions but recruiters and hiring managers as well sometimes so don’t get caught off guard.

Startups (Insights from Sajiv Shah - Intern at Figure and Relativity Space):

For this section, I talked to Sajiv Shah, a Umich MechE student who interned at Figure (humanoid robots) and Relativity Space (3D-printed rockets) and he gave some insight which I’ll detail below. Overall, expect your work to be heavily multidisciplinary. Design engineer serves as a catchall for many many things that you will be doing. For example, here is a video by Anduril of what a mechanical design engineer intern did during the summer. As you can see it not only involves designing a component, but then procuring the materials, manufacturing in-house (if needed), assembly, testing, and verification/validation.

How can you get a job like this? Well first of all, dive in headfirst to everything. What that means is to get as much experience owning mechanical assemblies that do things as you can. Join your FSAE team and become a system lead on the team (also make some life long memories — cannot stress this enough).

Start a complex pet project that involves numerous disciplines of mechanical engineering. Get involved in research projects that are very hands-on. In addition to this, track (and document) your projects rigorously and be able to justify the decisions and trade-offs you made as they will almost certainly come up in an interview. Be driven: prior to interviewing at Relativity, Sajiv took a gap year and spent significant time reading a materials and structures textbook as well as worked at a drone motor company which gave him an edge in recruiting. Any way you look at it, this is not a job you get with no experience (unless you get lucky or the startup is in a veeerrryyy early stage).

Getting an interview could be through a cold application, but you’ll have way more luck DMing someone high up in the company and telling them what you’ve accomplished. Checkout more on the Hardware is Hard LinkedIn Trojan Horse for this, but be very purposeful of who you are messaging and why. Although cold applying can work, I attribute the majority of my success to LinkedIn messaging. Now that LinkedIn has cut down how many people you can personalize your invitation to, I would honestly suggest getting premium and spending that $40 or so a month because it could pan out and turn into an opportunity that’s worth much more. This is something I see very few MechEs do. Here was an invitation I sent that directly led to a startup interview during full time recruiting:

“Hi ___,

I’m a Umich MechE MSE alum and I’ve applied to the team at ____! I’ve worked on challenging engineering projects at Tesla (in process development), Stellantis, Michelin (in quality) and more. I’d love to chat about how I could be a great fit on your team!

Same strategy as listed on hardwareishard.com and it’s short, descriptive, and to the point. This was for a manufacturing role so I highlighted the teams I worked on in manufacturing environments. Another thing I would suggest is follow startups and VCs that fund these companies on Twitter/LinkedIn and reach out to them there too. I specifically follow a lot of deep tech VCs to get this info (Cantos, Eclipse, Also Capital, Susa Ventures, etc.). The interviews themselves can be tough. A lot of times they will be multiple rounds and can include a project presentation, which is geared to see how you work through physics and design as well as how you can justify the decisions you made.

Now detailing a little more of Sajiv’s advice. At a startup, decisions are made very quickly and the product lifecycle is a lot smaller so be prepared to move fast. Teach yourself, as you will be learning a lot by doing. For example, on top of the design work, since Figure is a robotics startup, Sajiv has been teaching himself about controls and frequencies. In regards to getting an interview, be persistent in reaching out to people. It’s easier to get responses from smaller companies and startups to be honest. He heard of Figure through a reporter that he maintained a connection with. You never know where your next opportunity can come from!

Good luck in those midterms! Next up we’ll be breaking out some tips for manufacturing and nontraditional roles so stay tuned! 🎵