skip to content
Navid Mesbah
فا
← Back

AI Is Coming for Circuit Boards

/ 11 min read

Back to Where I Started

I studied electrical engineering at the University of Shiraz. I still remember the smell of solder in the electronics lab, the frustration of a circuit that worked in simulation and failed on the bench, and the quiet satisfaction of an oscilloscope finally showing the waveform you expected.

Later, I spent time in the smart energy management lab at the University of Tehran, working around smart grids. That lab showed me electronics at a completely different scale. A board is no longer an isolated object; it’s one node in a network of meters, sensors, controllers, and software that has to decide, continuously, how energy should flow. It taught me that hardware and intelligence belong together, and that the hardest problems live where circuits meet decisions.

For years after that, I drifted toward software, graphics, games, and AI. Electronics became something I did on weekends: small boards, sensors, the occasional project that went nowhere but taught me something. Software moved fast and offered more immediate rewards. Hardware was slow, expensive, and unforgiving.

But I never stopped noticing how far behind the tools are.

Software development has been transformed by AI in the last two years. Engineers describe what they want and agents write the code. Reviews are partly automated. Tests are generated. Documentation writes itself, more or less. The daily experience of building software in 2026 is very different from 2023.

Electronics design mostly hasn’t changed. You still spend hours searching datasheets for the right part. You still draw schematics largely by hand, connecting pins one by one. You still catch mistakes in review that a careful assistant would have flagged in seconds, if you catch them at all. And when you miss one, you find out weeks later, when the boards come back from the factory and don’t work.

That’s why I’m founding Zagros Forest. The same idea behind Charsook, humans and AI building together with the human’s judgment at the center, applied to hardware.

Why Hardware Matters Now

It’s worth stepping back to ask why this matters beyond the convenience of engineers.

The world is filling up with electronics. Not just phones and laptops, but sensors in farms, monitors in hospitals, controllers in factories, devices in homes, instruments in research labs, small machines in places that never had machines before. Almost every interesting idea in the physical world eventually needs a circuit board.

But designing electronics remains a specialized and expensive skill. A startup with a great idea for a medical device or an agricultural sensor needs electrical engineers, and good ones are scarce. A researcher who needs a custom instrument has to either learn electronics deeply or find someone who has. A student who wants to build something real often gives up when they reach the hardware.

If AI can lower the barrier to good electronics design the way it has lowered the barrier to software, the effect on what gets built in the physical world could be enormous. Not by removing engineers, but by making each engineer far more capable, and by letting people with real problems get further on their own before they need an expert.

There’s also a personal dimension. In regions where access to specialized engineering talent and fast prototyping services is limited, the gap between an idea and a working device is especially wide. I’ve seen many talented people with good ideas stall at exactly this point. Tools that carry expert knowledge can narrow that gap.

Why Hardware Is Harder

Hardware is less forgiving than software, and that shapes everything about how AI should be used for it.

A bug in code can be patched tonight. You push a fix, users update, and the problem is gone. A bug on a board means a new manufacturing run, weeks of waiting, and money you don’t get back. If the board is already in customers’ hands, it might mean a recall.

Electronics also involves physics in a way software rarely does. Current flows and creates heat. Signals interfere with each other. Traces have resistance, capacitance, and inductance. A design that is logically correct can still fail because two signals are too close together, because a power supply can’t respond fast enough, or because a component gets too hot. These failures are hard to see on a schematic and sometimes hard to see even in simulation.

And the information needed to design well is scattered across thousands of documents. Every component comes with a datasheet, often dozens of pages long, describing how it must be used: which capacitors it needs, how its pins should be connected, what layout rules apply, what happens at extreme temperatures. Application notes add more guidance. Manufacturers publish design rules. Experienced engineers carry much of this knowledge in their heads, built up over years of mistakes.

All of this makes electronics a harder problem for AI. It also makes it a more valuable one. Catching one serious error before fabrication can pay for a lot of tooling. Avoiding one failed board run can save weeks of a project’s schedule.

It also means verification matters even more than in software. Nobody should send a board to the factory because a model said it looked fine. Every AI contribution to a hardware design has to be checkable, and the checks have to be grounded in real rules and real physics, not in plausibility.

What We’re Building First

We’re starting with tools, not devices.

Text-to-schematic. Describe what you want: “a USB-C powered sensor board with an ESP32, a temperature and humidity sensor, a lithium battery charger, and a status LED.” Get a first-draft schematic with real parts, real footprints, and the supporting components that datasheets quietly require.

That last part is where most of the value lives. Anyone can place an ESP32 on a schematic. The hard part is knowing it needs specific decoupling capacitors near its power pins, a particular reset circuit, the right pull-up resistors on certain lines, an antenna clearance zone, and a power supply that can handle its current spikes. A good first draft gets these details right, citing where each requirement comes from, so the engineer can check rather than research.

The output is not a final design. It’s a strong starting point that an engineer reviews and refines, the same way a senior engineer might review a junior’s first draft. But it can turn days of setup into an afternoon of review.

Agent-driven PCB review. Agents that read your schematic and layout the way an experienced engineer would. They check decoupling placement, power integrity, thermal issues, footprint mismatches, missing pull-ups, incorrect pin assignments, clearance violations, and the hundred small things that go wrong at 2 a.m. before a deadline.

Design review is one of the best fits for AI in electronics. It’s exhaustive, rule-based in large part, and tedious for humans. A human reviewer, especially a tired one, will miss things. An agent can systematically check every component against its datasheet, every net against its expected behavior, every trace against the design rules. When it finds something, it explains what it found and why it matters, with a reference to the source.

The goal is not to replace human review but to change what human reviewers spend their time on. Instead of hunting for missing capacitors, they can think about architecture, tradeoffs, and the subtle issues that require real experience.

Placement and routing. Help with the slow, spatial part of layout: where components go on the board and how the copper connects them. This is where design becomes physical, and where much of an engineer’s time disappears.

Automated routing has existed for decades, but it has a mixed reputation. Traditional autorouters often produce layouts that technically connect everything but ignore the things experienced designers care about: signal integrity, manufacturability, thermal behavior, and simple readability. We think AI can do better by learning from how good designers actually work, and by keeping the engineer firmly in control of the decisions that matter. The engineer places the critical parts and sets the priorities; the tools handle the tedious remainder and explain their choices.

Why Agents Fit This Work

Electronics design is full of rules that live in documents: datasheets, application notes, manufacturer guidelines, industry standards. It’s tedious for humans to cross-check them and natural for agents.

An agent can read a datasheet for every component on a board, extract the requirements, and check each one against the design. It can do this for every component, every time, without getting tired or skipping the boring ones. That’s exactly the kind of work that humans do poorly and machines do well.

The engineer’s real value is in tradeoffs and judgment. Should this design prioritize battery life or performance? Is this component worth its higher cost because of its better availability? Should we add a test point here for manufacturing, at the cost of board space? Is this design robust enough for the environment it will live in? These are questions that require understanding the product, the users, the business, and the physical world. They’re not rules to look up.

Remembering which pin needs a 100nF capacitor is not where human brilliance should be spent.

So the pattern is the same one I use in software: the human frames the problem and makes the calls, agents do the exhaustive checking and drafting, and everything gets verified before it becomes real. In hardware, “before it becomes real” has a very concrete meaning: before the files go to the factory.

Lessons From Software

Building AI tools for software at Charsook taught me several things that I’m applying directly to hardware.

Agents need context. An agent that doesn’t understand the purpose of a design will make plausible but wrong choices. In software, this means giving agents good specifications. In hardware, it means giving them requirements: operating environment, power budget, cost targets, manufacturing constraints. A schematic for a battery-powered outdoor sensor and a schematic for a lab instrument should look different even if they use similar parts.

Evals are essential. We’re building eval sets of real designs, including known mistakes, so we can measure whether our tools actually catch the errors they claim to catch. A review tool that finds nothing is useless; a review tool that flags everything is equally useless. Evals tell us where we are between those extremes.

Trust is earned per claim. Every finding from a review agent should come with evidence: the datasheet section, the rule, the calculation. Engineers shouldn’t have to trust the agent; they should be able to check it. This is the same principle behind my interest in formal verification. Shrink what needs to be trusted until a human can verify it.

Respect existing workflows. Electronics engineers have tools they’ve used for years. A new tool that requires abandoning everything familiar won’t be adopted, no matter how clever. We’re designing to work alongside existing tools and formats, not to replace them overnight.

Then, Devices

Once the tools are good enough that we trust them on our own work, we’ll build our own devices with them.

That’s the honest test of any tool: would you bet your own product on it? If our schematic generator, review agents, and layout tools are good enough that we’re comfortable sending our own boards to fabrication based on them, then they’re good enough for others. If not, we have more work to do.

I don’t want to say too much yet about which devices. But I’ll say this: the ideas come from the same place as everything else I build. Creative tools, education, and instruments for understanding the world. Physical objects that help people make, learn, and see.

A Long Game

Hardware is a long game. Software companies can iterate daily; hardware companies iterate in months. The tools we’re building now will take time to mature, and earning the trust of electrical engineers, who are rightly skeptical of anything that claims to make their work easier, will take longer.

I’m comfortable with that. Some of the most meaningful work takes the longest. And there’s something deeply satisfying, after years in software, about returning to the field I first studied, with tools I couldn’t have imagined back then: in the electronics lab at Shiraz, where the smell of solder first told me I was making something real, or in the smart grid lab in Tehran, where I first saw circuits and intelligence working as one system.

I’m excited to find out how far we can go.