A Week of PCB Layout in an Hour: Quilter Trains Agents, Not Imitators

Will Summerlin opens Autopilot by congratulating Sergiy Nesterenko on a $10 million round led by Benchmark, then asks for the founding story. Nesterenko's first job out of college was at SpaceX, where he owned Falcon 9 and Falcon Heavy second-stage avionics in high-radiation environments. In his own shorthand, the job was to make sure the rocket did not die when it went to space. That work meant living inside electronics, and living inside the slow, physical loop of developing hardware and all the hardware that supports the hardware.

The cultural export he still cites almost daily is having something exciting to work on. At SpaceX you walked in the door knowing the company was actually trying to go to Mars, and that every decision had to point that way. Recruiting got easier. Showing up got more fun. People worked hard on the current project to earn the next cool one. He has tried to import that missionary energy into Quilter, where the team, he says, still arrives with a lot of it.

Quilter did not start as a circuit-board company. In 2019, in a different capital market, he met investor Michael Dearing, then at Harrison Metal, and began talking about what a company might even be. The first thesis was low-cost sensors for manufacturing. A few months later he decided they were too late. Analysis on those sensors was next, then reliability analysis on circuit boards. The turn that stuck was almost a tautology: why analyze a board after the fact if you can design it correctly in the first place? It took nearly a year of wandering before the company landed on creating circuit boards.

Asked to explain circuit boards at a fifth-grade level, Nesterenko holds up examples. A board is a rigid piece of fiberglass and copper that connects chips. The chips compute, store, and supply power; the board is a passive glue that lets them talk. Open a computer, a TV, a light bulb, or the control panel of a stove and you will find one. Almost every electronic product gets its own board. Creating a new chip is a multi-million-dollar proposition, so chips are reused. Boards are the opposite: you take existing chips, glue them in a new way, and make a new product. The iPhone board is not the Android board, and neither is the one in a laptop. Difficulty scales with two things: combinatorial density, how many parts are packed how tightly, and physics, high-speed signals and high-power heat. An Nvidia video card and an alarm clock are night and day.

He maps the design process onto software because more people have written code than laid out a board. First you draw a schematic, a human-readable block diagram of inputs, outputs, and function, closer to writing Python. Then you convert that schematic into a layout, the blueprint that tells a factory how to arrange copper, fiberglass, solder, and components, closer to compiling. Then you send the blueprint out, get a board on your desk, discover you messed something up, and start over. Layout time ranges from a couple of hours for a handful of parts to two or three months for a Starlink antenna or an iPhone motherboard. The irony is that manufacturing is often the fast part. Pay for it and a prototype can be on your desk in 24 hours, or a few days from China. The wait is the human compile.

Basic manufacturability is already handled. Design rule checks run quickly in ordinary tools: nothing collides, every net matches the schematic, copper is far enough apart that etching will not bleed. The hard part is physics that the DRC does not see. Crosstalk is his example. Two traces that never touch can still couple energy, because the signal is an electromagnetic wave. In the extreme case, more energy lands on the victim than on the intended wire, and the board simply does not work. A designer may have hundreds or thousands of such possible imperfections. People use rules of thumb and some targeted simulation. There is no one-click thumbs-up. Practically speaking, he says, it almost never works the first time. A design poor enough to radiate will also fail FCC testing in the United States and cannot be sold.

The principle is to eliminate layout the way compilers eliminated hand-written machine code: click a button and it runs. That is much easier said than done. For boards inside today's scope, the designer finishes the schematic, opens a board file to pin down hard physical constraints such as outline, connectors, buttons, and a screen, leaves the rest empty, uploads it, and Quilter drives it to something a manufacturer can build. Give it a computer motherboard or an Nvidia card today and you will not get a good result, and probably not a completed one. Even "easy" boards whose physics is not that complicated can still take a week or two by hand, so there is already a market in killing that latency. The longer ambition is boards better than humans can design.

PCB layout is its own profession. He puts the U.S. median salary around $100,000 to $110,000 a year, but the labor is not only those specialists. Almost anyone who designs electronics eventually lays out boards; the professionals are reserved for complicated, mission-critical work. SpaceX had a few dozen layout people, enough for flight computers and intense radios. In a role like his, radiation testing, you did your own designs because the central team could not support you, and it could easily take half your time. Behind every flight computer sit tens or hundreds of early prototypes, plus fixtures to test memory chips, power chips, and support equipment. He estimates $20 to $50 billion a year of labor on the task, and a drastic undersupply. Young engineers, in his view, are choosing software, AI, and robotics; fewer are going into electrical engineering and intense board design. Even a five- or ten-person IoT shop still needs someone on the electronics, whether they layout in-house or pay an outsider.

What a hardware company cares about depends on who you are and what you are building. For an early prototype, speed dominates: the fastest path to a board on the desk. For Apple laying out an iPhone motherboard, size and performance come first, with cost in the mix but not always the thing you squeeze hardest. A well-designed board, in his simplest test, faithfully recreates the schematic. A poorly designed one keeps the chip and memory from running at the speed they could. Slow everything down and it works; turn the communication up and it fails. You do not want to be bottlenecked by fiberglass and copper. On the current arc from "as good as an average designer" to "better than the best," he is frank. Quilter is not yet as good as the average designer on physics and density. The only place it is significantly better is time: on boards where physics is not the constraint, an hour of Quilter against a week of a person. He thinks catching and then passing most humans on the physics is achievable within five years.

If the system ever beats humans on physics, the first dividend is cost and size. A designer staring at two or three months of layout is terrified of discovering, a few weeks from the end, that the design will not close. Resizing the board or adding layers can add another month, which is unacceptable. So people make conservative assumptions: boards a bit bigger than they need, extra layers, thicker traces, more copper pours. The result is larger boards, lower manufacturing yield, and a higher bill. A machine that is not afraid to fail a layout in simulation can explore more aggressive packs and take that waste out.

The stack is not "GPT-4 for boards." The analogy he wants is DeepMind and AlphaGo. Reinforcement learning means a game, an agent that plays it, and a score. DeepMind's game was Go. Quilter's game is circuit-board design; the score is physics and manufacturability: how cheap is it, and will it actually work? One virtue of that setup is that you do not need a corpus of human boards. Plenty of people have tried the obvious alternative: show a neural net lots of human layouts and ask it to recreate them. That path is blocked twice. There is no GitHub of millions of great designs, and the best ones are locked at Nvidia, Apple, and SpaceX. More important, learning from human examples is likely to peak at roughly human performance. That is what happened when DeepMind first trained on human Go games, queried the net for the most human-like move, and reached grandmaster level without becoming world champion. The winning strategies came from the network trying new things, seeing what happened, and remembering them.

Summerlin notes the unexplored solution space outside the boards humans have already played. Nesterenko agrees that is where superhuman layouts would come from. The same logic, later in the conversation, is why he wants the board right on the first try. In principle, whether a board will work is computable: capture the manufacturing process, Maxwell's equations, and the laws of thermodynamics. It is a lot of compute and a lot of modeling, but it is possible, and it is worth doing. In software you sprinkle print statements and rerun. On a board, finding what you did wrong can take weeks or months, and some boards ship with the issue never figured out.

Compounding in this business is not only "more proprietary data." Inside one customer, boards start to look alike. SpaceX forced engineers onto a small, screened component library for supply-chain and flight-worthiness reasons, so a new board reached for a part you had already used rather than ten market alternatives. Even an individual designer reaches for the chip already in the toolbox, the way a programmer reaches for a familiar package. That opens the door to little per-customer networks that stay on the customer's data and get special precision on the parts they actually ask for. Quilter does not do this yet. He has thought about it.

The other scale loop is manufacturers. Today you often have to pick the fab first and design to its rules. In software you write the code and choose the compile target last. As Quilter integrates more factories, it can serve them more designs, help their margins, bring cost down, and let the customer submit without that upfront marriage. By volume and by unique designs, China still makes the most boards. People still fab in the United States for two hard reasons. One is regulation: SpaceX cannot send designs to China because of ITAR. The other is latency. A U.S. shop can build a board and drive it over the same day; China is bounded by shipping. The biggest fab in China, he says, processes about 8,000 unique designs per day. Simple, thick-trace boards are relatively commoditized. Differentiation shows up in high-speed materials, tolerances, yield, quick-turn versus high volume.

Scaling laws are not the same as in language models. Bigger models and more compute help, but they are not the only bottleneck. Reinforcement learning has three pieces, the game, the agent, and the score, and a huge agent is useless if the game is slow to evaluate or the score does not match real life. They invest across all three, and they avoid problems that need a million GPUs for three months, which are expensive and slow to iterate. His favorite demo is a robot that learns to walk in simulation in five minutes on a single GPU. Using Google's JAX library, the physics engine itself runs on the GPU, so you can evaluate on the order of 50 million episodes, about a thousand steps each, in a few minutes. Fast iteration, he argues, beats hoping a giant model is the right one.

Outside PCBs, the nearest cousins are chiplet design and chip design, still electronics. He would also like to see mechanical versions of the same method: the right way to arrange fins on a heat sink, how to design a fan, even a wing with less turbulence. He is not an expert in those domains. The pattern Summerlin names is design optimization with well-defined outer constraints, the laws of physics, and a wide interior to explore. Nesterenko agrees. Wherever you can define a score, the method has a chance.

Go-to-market is gated by what the software can actually do. In a five-year future where it can do a computer motherboard, you walk into a head of electrical engineering and show millions of dollars of value immediately. They cannot do that yet. The segment that is working is what he calls electrical-engineering-adjacent roles: people who need a board to do their job but whose job is not the board. His SpaceX role was exactly that. Firmware engineers, robotics engineers, automation engineers, and test engineers show up. They are not designing the main Apple motherboard. They are designing the things next to it. Interest is less a function of vertical, aerospace versus consumer versus chipmaker, than of who you are talking to inside the company. Every one of those firms has intern-easy boards and boards that are nearly impossible. The companies pulling hardest right now are the ones for whom a board on the desk tomorrow, rather than next week, is critical.

Pricing is meant to tap a budget that already exists. If you outsource layout today, shops look at pin or connection count and charge somewhere between $1 and $10 per pin, depending on country and difficulty. Quilter plans to enter on the cheaper end of that range and move up as simulation and value improve. The product is in open beta at quilter.ai. Electrical engineers can upload a design and try it. His slightly ambitious internal goal is to take a video call on a Quilter-designed computer by the end of the year, likely something Raspberry Pi-shaped rather than a full laptop, and he offers to send Summerlin one if they pull it off.

Hiring has been easier than the generic AI labor market would suggest. Reinforcement learning is a rare industrial application, maybe 10 to 100 times less common than supervised learning, so RL experts are drawn to a place that actually uses their work. The other magnet is physicality. This is not a slightly better chat satisfaction score. It is something you can hold. On the worry that talent is disappearing into closed labs and taking the research with it, he is unconvinced. Historically, when companies vacuum up talent and money, you still get open-source alternatives and a busy ecosystem. Even if the best RL engineers go to DeepMind and DeepMind closes everything, people will still study the field and build open projects in order to get hired there.

On open versus closed foundation models over the next five years, he refuses a strong bet. Open source has been incredible and underestimated: Windows still owns the home desktop, Linux runs production servers despite Microsoft's efforts. Neural nets may be different. Open source shines when a critical mass of small contributions makes a great product, as with 10,000 people writing VS Code extensions. It is not clear that a thousand people each spending an hour a week produce a better neural network. That is an argument against open source being quite as competitive as it used to be. It is not, he says, a prediction.

Asked what he would tell himself on day one, he does not pretend he could skip the wilderness year. The useful investment is accounting for how difficult, lonely, and brutal a startup is. Surround yourself with great people: a co-founder if you have one, friends building hard things, investors you evaluate as relationships ahead of terms, because most founders just quit and surviving is the job. The other half is self-care. A startup is more stress and self-criticism, on the whole, than working at a place like SpaceX, even if a launch day is worse than an average founder day. Routine, family time, exercise, food, and sleep are obvious and easy to forget once the work swallows you. He chose Benchmark after an introduction from an existing investor and diligence on partner Eric Vishria. Experience seeing companies from seed to public was attractive. What closed it was the references: he had never heard founder calls that positive, bar none. Summerlin admits that in a previous startup he went six months without a couch until his girlfriend told him to buy one.