Quilter: Shaping the Future of AI-Powered Circuit Board Layout

Printed circuit boards are the backbone of modern electronics, integrating with nearly every device that beeps, clicks, or flashes in our daily lives—from the smartphones in our pockets to the refrigerators that keep our food fresh.

For decades, people have tried to automate the design of circuit board layouts, but despite automating parts of the process, it largely remains as meticulous and time-consuming as manually compiling code into a machine-readable format.

Quilter leverages AI to automate the design of circuit board layouts and accelerate innovation in electronics. Founded by Sergiy Nesterenko, the startup recently raised $10M in Series A funding by Benchmark with participation from Coatue and existing investors Root Ventures and Harrison Metal Capital.

Learn more about the future of AI-powered circuit board layout from our interview with the CEO and founder, Sergiy Nesterenko:

Why Did You Start Quilter?

Even before I had a concrete idea, I always wanted to found a startup. After my studies, I started working at SpaceX, learning the ins and outs of designing electronics in real life. It seemed puzzling that engineers still spent dozens of hours designing the layout of a circuit board manually—why wasn’t a computer doing that? The pain was obvious to me, and I thought there could be a way to solve it. 

I never imagined myself working at one company for decades, and when I felt I had made a good contribution at SpaceX, after five years, I moved on. I spent four months working for a startup as a contractor and then helping an investor evaluate another startup. When we finished the due diligence, he turned around and asked me whether I wanted to found a startup myself. I did some soul-searching, but that was the final trigger. It flipped the switch and led me to found Quilter to automate the grunt work around designing circuit board layouts.

How Can AI Help With Circuit Board Layout?

Let’s start by explaining how designing circuit board layouts works today and drawing some analogies to software development, which most people are more familiar with. 

It generally involves three steps. The first step is designing a circuit schematic, a flowchart, or block diagram that defines the functionality of the circuit board in a human-readable, high-level way. It’s like writing code to define a program. 

Next, this high-level schematic is implemented by a set of actual electronic components in a blueprint called layout, defining where inputs, outputs, connectors, and other components go to faithfully recreate the circuit board’s intended functionality. It’s like compiling code to a machine-readable format—and it’s as hard as compiling code by hand. Electronic engineers currently spend months drawing traces using programs not that much more sophisticated than Microsoft Paint. 

As a final step, the blueprint goes to a manufacturer, who builds an actual circuit board. It comes back for testing—and most likely, it doesn’t work out of the box. That’s why it usually takes several iterations to produce a working circuit board layout. 

We want to automate the second step, making circuit layout completely autonomous, and let electronic engineers focus on the first step, writing high-level schematics. We focus on one job function, solving one entire problem for electronic engineers while plugging neatly into their existing workflow. 

We leave the component selection to the engineers—large companies even have librarians to manage component libraries. So, you upload a schematic and component library, and we can account for all kinds of inputs and constraints. Maybe you need a specific form factor, where you want a specific outline of your circuit board or specific locations for input, outputs, and connectors. Given all the inputs, we can place and route all the other components automatically. 

While people tried to automate circuit board design, it turned out to be a notoriously tricky problem to get right, as imperfections or cross-talk can easily impair signals. Sure, it takes more to disturb a digital signal and literally flip a 0 for a 1, whereas any disturbance affects the absolute value of an analog signal. But even digital design remains largely manual, even though people have been researching solutions for about 60 years. Some have tried to tackle it as a graph embedding problem, but that didn’t work, as in a circuit board, the nodes and edges are not infinitely small.

That’s why, thus far, automation has been limited to small subsections of a circuit board and low numbers of connections. Most current algorithms simply don’t finish the job. And when you get a 90%-complete design, it’s really hard to make it work by hand, so you might as well start from scratch.  

We’re using AI to automate circuit board layout. While large language models get a lot of the hype currently, we only use them to extract information from datasheets of electronic components. Instead, our main focus is on reinforcement learning. It’s more like an AI agent playing a game where it lays out a circuit board. We define a reward function that tells the AI agent whether it’s playing well, and we define what actions it can choose in the game to make sure the resulting circuit board will work. 

There are things we can validate at runtime, e.g., check for collisions between different components, which gets integrated into the reward function. I.e., if two components collide, the AI agent gets a lower reward and will avoid it in the future. Other things are too expensive to evaluate at runtime, e.g., cross-talk, so we create a score after the game has been played, i.e., the layout of a circuit board is finished, and we train the AI agent to optimize the score and create increasingly better layouts. The AI agent learns circuit board layout from the ground up—without data input, just from the physics.

One challenge is to plan ahead, i.e., place components and traces in a way that allows you to place some more components after the first 10,000 components have been placed. Humans are surprisingly good at this high-level spatial reasoning, and we needed to teach the AI agent the same capability, using planning algorithms or by making the game easier initially by letting it move around traces even if they have already been placed on the board. Over time, this will allow the AI agent to tackle increasingly complex circuit board layouts. 

How Did You Evaluate Your Startup Idea?

The electronics market is huge, so the main question is how to capture it as an up-and-coming startup. We do that by having a differentiated product that is much better than manually creating layouts. 

It’s good to remember the sheer amount of time and effort that is currently spent on layout, requiring dozens of people and hundreds of hours. That’s a lot of time, and it’s already budgeted for. Before Quilter, you emailed a firm specialized in layouts, and they charge you per pin. Now, you drag and drop your project into our website, and it runs overnight to produce a layout. We integrate smoothly into an existing workflow and take a task that took weeks and complete it in hours. For a large company like SpaceX, this also means faster time to market and less overhead. 

Quilter has been in closed beta with a small community of engineers who have generated over 100K unique board layouts. We’re now in open beta, and everyone is welcome to try it and play with our tool—you can sign up for the beta here: blog.quilter.ai/open-beta*

What Advice Would You Give Fellow Deep Tech Founders?

Entrepreneurship is 100x harder and far less glorious than all the hype it gets in TV shows. None of that is important. Surround yourself with people who will be there for your entrepreneurship journey. Be prepared to be punched in the face and get up again. Many startups fail because founders run out of steam. 

Want to Know More?

Learn more about how you can save valuable engineering time on our website at www.quilter.ai*. 

You will find more insights into our technology, how we automate designing circuit board layouts, and our ‘Fab for Free’ program on our blog at blog.quilter.ai*

*Sponsored links—we greatly appreciate the support by Quilter

Comments are closed.