Judy Warner (00:01.48) Hi, Sergi. So good to see you again. And I'm looking forward to hearing what's going on at Quilter. So before we really dig into the good stuff, why don't you take a moment and introduce yourself and tell us a little bit about Quilter. Sergiy Nesterenko (00:16.194) All right, Judy, good to see you again. Hello, everyone. My name is Sergei, CEO at Quilter. At Quilter, we're trying to automate PCB layout. Judy Warner (00:26.704) Really? That's all you're gonna say? It's like, I'm gonna build the Empire State Building. Yeah, okay. All right, we'll save the good stuff for later, Sergi. So, well, I wanna go back, because you have an interesting story of how you ended up being the CEO of Quilter. So I wanna capture a little bit of your cool background. So let's talk about your background at SpaceX. Sergiy Nesterenko (00:28.366) we've got a lot more than that, but... You Judy Warner (00:54.244) And how that sort of, that was a formative time and sort of how that culture of first principles really impacted you before you started Quilter. Sergiy Nesterenko (01:05.006) Yeah, yeah, thank you. So I think SpaceX is awesome company, obviously. I had spent about five years there in the avionics group. In particular, I was responsible for Falcon 9 Falcon Heavy's second stage in high radiation environments, which means when the second stage makes orbit, it's going to get bombarded by particles. Particles tend to break, ionized particles, they tend to break electronics. My job was to make sure that that didn't actually compromise the mission. And I think, you know, the cool thing about SpaceX is that like there's one super long term vision, and it's a really hard problem. And everybody there is just motivated by that one thing. And we're all trying to solve it, right. And so, you know, was very kind of low bureaucracy, low ego, very high, like, how do we just get it done? It's super hard, we'll figure it out kind of culture, which was awesome. And, you know, for me, it's just an excellent place to grow up as an engineer, right, you get Judy Warner (01:42.247) Right. Sergiy Nesterenko (01:59.096) dozens of projects every year. All the projects are meaningful. Nobody has time to do kind of meaningless projects. And so it's a real trial by fire. And I think that's really excellent for young engineer. Judy Warner (02:10.214) I think so too. And I think I shared with you for a short period of time, I was selling RF boards to SpaceX. So got to walk through there three times, which was just amazing. But you can really feel that what a friend of mine, Jeremy Blum calls GSD. We'll call it get stuff done, but you get my drift. It's that. It's like, well, you can't get it. We'll just kick the wall down then. Sergiy Nesterenko (02:30.68) Yep. Judy Warner (02:37.488) make a new, it's just there is such a, and things get done, you know, and it really is a neat environment to be in. And I've never in all my years experienced a culture like that in a company. So. Sergiy Nesterenko (02:54.414) Yeah, you know, the biggest compliment we received, at least to me, obviously, we worked very closely at that. So this was like 2013, 2014, we were working closely with NASA to start working on, you know, supplying the space station astronauts, all that kind of stuff. And we had a lot of the NASA folks, you know, reviewing our work, making sure everything's getting done well, you know, certification, all that. And some of those guys were from the early shuttle or even Apollo era. And some of them communicated that like that, that feeling that we captured that that mentality of the engineers. Judy Warner (03:22.27) Mm-hmm. Sergiy Nesterenko (03:24.012) resembled the Apollo, the kind of like, we've got to do this, it's short. like, and they'd been reminiscent that NASA had lost it a little bit. And it was great to see it back somewhere, right? And so that was probably the biggest compliment, because I mean, as a kid, I loved Apollo, and those were heroes of mine, right? So to hear that we might have gotten a taste of what that might have been like was a huge compliment. Judy Warner (03:26.966) okay. Judy Warner (03:31.986) Yeah. Yeah. Judy Warner (03:38.289) Right. Judy Warner (03:44.978) Wow, that's fascinating. I love that. So I know that you, Judy Warner (03:53.734) You tell us a little bit define what you mean when you say first principles and sort of, I mean, that's sort of how you encapsulate that culture. So tell us a little bit about how you would define that and then how that started sort of seeding the idea of what you're doing now. Sergiy Nesterenko (04:11.406) Yeah, sure. mean, so first principles, you I guess my definition would be, you know, to, uh, the simplest way to describe it, maybe it's just to avoid thinking by analogy, right? Um, I think that, um, you know, you can't, um, when you tackle a new problem or you tackle a problem in general, right? Like you can't read, arrive everything from scratch, right? If I ask you to like, if ask one of my engineers to, go to a website and make a new button, they sure shouldn't be redesigning the processor and the compiler and the operating system. And HTML and CSS and JavaScript, TypeScript, like those abstractions exist for a reason, we don't need to dig into those, right? And if we tackle them every time, that would be bad. And so this gives rise to this kind of building on top of abstractions, reusing previous decisions and thinking by analogy, which are frankly almost always good things. Sometimes they're not, right? Sometimes what you'll see is that in an industry or in a project or in a company, know, somebody will make one decision, Judy Warner (05:00.497) Right. Sergiy Nesterenko (05:10.796) And then that decision gets reused for 10, 20, 30, 40 years. And oftentimes the conditions that we have now drift from the conditions that existed when the decision was initially made. Right. And so what's really important is to recognize when that has happened. And at that point to go and question why the decision was made and potentially make a different one based on the new information. So I think of first principles thinking really as most importantly, identifying when that has happened, identifying when it's important to stop. Judy Warner (05:18.885) Indeed. Judy Warner (05:29.479) Mm-hmm. Sergiy Nesterenko (05:39.631) question the decision's been made and go all the way back to why it's been made, right? And a lot of times, especially in a difficult industry like aerospace, you are forced to do this, right? You're forced to. A really obvious example is like, you know, at SpaceX, we initially were like, okay, we need TPUs to fly the rocket. Okay, what is the industry use? Industry uses rad hard parts. These are parts that are designed not to fail in radiation. Okay, that's an obvious default choice to just proceed with that, but they're really expensive. And so we can't very, very much for listeners who don't know, we're talking about like a half a million dollars per chip kind of expensive. And this is like a 200 megahertz, know, 200 megabytes RAM kind of computer, right? We're not talking super computers here. So at that point, you're forced to reassess like, man, do we really need that? Does it need to be that expensive? Do we need it to be right hard? Can we do something else? Can we put three of them and have them be redundant? Can we reboot them? And so you're forced into that decision. Judy Warner (06:07.954) very. Judy Warner (06:14.481) Yeah. Sergiy Nesterenko (06:33.134) And so sometimes you're forced into it, which is great. Sometimes you're not, in which case it requires frankly, open-minded people to question those things, right? So SpaceX was obviously very good at that, right? We had to question everything for first principles almost by necessity, right? Because like what we were doing was perceived as impossible at the time. And it was because all of the previous decisions had led to that conclusion. And it causes you to think about in general what other things can be improved. Judy Warner (06:41.819) Mm-hmm. Judy Warner (06:48.86) Right. Judy Warner (06:52.957) course. Right. Sergiy Nesterenko (07:02.03) So for us, obviously relating to the audience here and to this podcast, a big bottleneck was circuit board design. We'd run into this problem where, okay, you're designing the rocket, you're designing the engine, you know what fluids are gonna flow in it, you what sensors you need for it, actuators. Then you start thinking about your schematic and then drawing it out and you can put a bunch of people and get in the schematic done quickly. But then we get stuck with one layout person, drawing trace by trace the whole board. Judy Warner (07:27.997) Yeah. Sergiy Nesterenko (07:29.71) you know, for a month at a time. And that means they're in critical path. And if they're in critical path, that means they're not just delaying the board, they're delaying the whole company. And so obviously we always questioned like, is that? Right. And even at that point, I was thinking like, why isn't this automated? And then, and then I try auto routers and realized why. Judy Warner (07:48.83) Yeah, those old stinky auto routers. So you've brought us to a perfect point of, okay, so you found that problem, which by the way, we've talked about a lot here. and I just have to say, I love what you are doing. And I love what the, I love how many for me, younger engineers are out there. Like this is just dumb. Like. They don't do this with semiconductors. Why are we doing it this way? And I love that you're all tackling this problem and it's a very interesting time and I'm tracking you so closely because I think it's wonderful. So let's talk about how you jumped from, this is a huge bottleneck. How do I solve this problem? Which by the way, as much as I say it, it's not an easy problem to solve. So. That's sort of a little bit about how you caught the vision, but what made you make that jump and say, I'm going to solve this problem and here's how I'm going to do it. Sergiy Nesterenko (08:54.37) Yeah, I mean, to add to what you're saying about it being difficult, you when I was doing initial research in this, found that like academic papers tackling specifically place and route for specifically circuit boards go back to the early sixties, right? So they predate CAD. Yeah. So it's like, it's 60 years of research now and here we are. So, so lots of respect for the problem and for everybody who's tackled it, right? It is very hard. Judy Warner (09:08.61) Why? I know. Judy Warner (09:15.515) Alright. Sergiy Nesterenko (09:21.846) And I think humans, I think people who have done it appreciate why it's hard. A lot of times I talk to like software engineers who haven't done PCB design and they're like, wait, that's still manual? And you have to kind of explain, but very hard problem. So I think for me, there were two things that came together as a like, okay, we should do this, right? From first principles. Thing number one is this is an obvious problem, right? Like it is very difficult. is... Judy Warner (09:25.734) Indeed. Judy Warner (09:47.399) glaring. Sergiy Nesterenko (09:49.08) Yeah, glaring, that's a great way to put it, right? Like it's slowing down companies like SpaceX, but also everybody else, right? Like it doesn't matter what big tech hardware company you name, all of them share this problem. You know, and it specifically sits at this critical path point where you're like done with a lot of your design and you're trying to get it out and test it and build it. And you can't get to it because of the layout. So the kind of the problem is, as you put it, glare. Judy Warner (09:59.752) Mm-hmm. Sergiy Nesterenko (10:14.009) That's very important. And it's very important when you start a business, when you pick a project, when you pick a PhD thesis, to do something that really tackles something difficult, something important. And so this checks that box very obviously. I think for me, the second moment is this question of why now? And given that people have been trying for 60 years, lots of very, very smart people have been trying for 60 years, what's changed that makes it possible when maybe it wasn't possible before? Judy Warner (10:26.268) Right. Judy Warner (10:41.297) Mm-hmm. Sergiy Nesterenko (10:41.711) And I think it was a couple of things. The first and biggest one for me was specifically AlphaGo. So this was a project led by DeepMind back in 2017 that beat the world champion at this game of Go. And for 20 years, it had been thought that, okay, we solved chess, but Go would be impossible because it's so much more combinatorially difficult. And Go approached this problem with reinforcement learning and neural networks. Judy Warner (10:58.461) Mm. Sergiy Nesterenko (11:09.903) and basically taught itself how to play the game without human supervision. Eventually there's detailed service, but eventually, and to me, this was a realization that like, okay, it is in general possible to give a computer a game and a clear win loss criteria and to have it teach itself. That's magical, right? That means that like, if you can define actions that a computer can take and you can define the success criteria, there is some hope that it can eventually meet and beat humans at that task. Judy Warner (11:39.121) Right, right. Sergiy Nesterenko (11:40.143) And so for me, the importance of the problem in the existence of technology that has those characteristics are, you know, one plus one equals two. Like that's, that means that we should go and try to apply that to this. Judy Warner (11:46.351) Mm-hmm. Judy Warner (11:52.914) So backing up just a second, I love what you're saying, by the way. I think it's really great that you're able to identify the specific things that are happening. And certainly there's so many technologies available today that weren't back in the sixties. And how we leverage that is, and now of course AI is flying and who knows where we're going to end up. Sergiy Nesterenko (12:09.817) Yeah. Yeah. Judy Warner (12:20.317) I want to go back to your engineering discipline. So you talked about a PhD. So I'm going to read here. You triple majored in math, physics, and chemistry at Berkeley. Why? Why those disciplines? Sergiy Nesterenko (12:36.271) Why? Yeah, yeah, that's a great question. So before I got to Berkeley, I'd obviously taken a lot of math, physics, chemistry classes. And when I got to Berkeley, I started to be faced with the decision of choosing my major. And in particular, what I was thinking at the time is that I would go and get a PhD, right? I thought I was headed into research, academia. My kind of perceived trajectory was that you have to go get a PhD. Judy Warner (12:51.047) Mm-hmm. Sergiy Nesterenko (13:01.935) Learn how to research learn how to invent something once you have an invention you can productize and start a company Turns out that's not the only way to go do that But nevertheless that was the thinking at the time and what I'd realized when it started Cal is that I knew enough about math physics and chemistry to know that I know nothing about math physics nor chemistry And and as I was thinking about that, I'm like man I'm choosing like my life's trajectory here based on such incomplete information like that seems silly and also like math physics chemistry is the foundation of everything else, right? You get into computer science, lot of it is math, mathematicians from the 50s, 60s, 70s that built that up. You know, a lot of engineering is derived from physics, right? So if you study statics and dynamics and physics, you can apply it to mechanical engineering, electrical engineering, so on and so forth. And chemistry likewise, right? With biology, biotechnology, chemical processes, chemical engineering, on and so forth. So I thought that like, Okay, I should just learn the fundamentals, right? And then if I do research in one of them, that's great. And if I end up just going into engineering, like I'll have the foundations to learn whatever I need in engineering. And so I gave myself the challenge to just go in and earn all three majors. Judy Warner (14:06.951) Well, apparently you're a little bit bright, Sergey. Sergey, sorry. I keep mispronouncing your name. I'm going to redo that because I have something in my throat. second. Sergiy Nesterenko (14:18.595) No worries. Judy Warner (14:23.384) Ahem. Judy Warner (14:29.06) I don't want to cough in here. Well, apparently, no, I'll do that again. So your cups not up. Well, apparently, you're an intelligent young man, uh, Sergey. So, uh, but I love the way, um, that you're explaining how you think. Right. Um, and that you're, it seems to me that you're wired at, um, looking past the obvious, right? That. Like you said, is math finite? Are any of these things finite? No, we know what we know today, but that's always evolving. And so I think it's so interesting to hear about the way that you think. So let's get back to automation. You mentioned auto routers. I actually made a podcast with somebody. think maybe it was, it was probably Duncan Haldane who you know. And I just couldn't think of anything better, but it's like an auto router that doesn't suck. You know, it's because everybody says auto router suck and I don't know a nicer way to say that. But you've made a case that why they suck, right? And again, this is about the way you think. So can you talk a little bit about why you think auto routers fail and why your approach at Quilter is different than what you've seen in the marketplace? Sergiy Nesterenko (16:00.814) Awesome. Yeah, yeah, happy to. So let me kind of walk through with a bit of an analogy of what I think the final stage should be, why they haven't met that yet, and then of course how we're addressing that, right? So, you know, from first principles, when you think about it, a PCB layout is really a constrained optimization problem, right? Like you're given a schematic and you are trying to make a blueprint effectively for a factory that makes that schematic real. And you have a couple of objectives, right? First objective is the board's gotta work, right? Very important, the board's actually gotta work. From there, there's objectives around yield, manufacturability, reliability, potentially cost, simplicity, build materials, those kinds of things, right? But at end of the day, it's a very well-defined objective. And so one of the kind of analogies I always come back to that kind of exemplifies what I wish this was is what's happened in software. So if you think about it, a schematic design is a little bit like writing code. You are expressing design intent. You're stating, here are our inputs, here are our outputs, here is the function we're going to perform in the middle. And it is primarily meant for consumption by other humans. Like Python is not meant to be read by a computer, it's meant to be read by a human. A schematic is not meant to be read by a computer, it's meant to be read by a human. It's humans communicating about design. Judy Warner (17:25.213) Mm-hmm. Sergiy Nesterenko (17:25.552) But of course, at some point we have to make like the atoms of the world dance, right? And so in software, you get a compiler that interprets that human text and turns it into machine instructions, right? Into additions and memory reads and memory writes and all that kind of stuff. And thankfully we're at the point today where you ask a software engineer when was the last time they looked at the machine code coming out of the compiler, right? Like most people laugh at that when I ask them that in software. Judy Warner (17:38.717) Mm-hmm. Sergiy Nesterenko (17:52.805) Which is great. And it didn't used to be the case, right? In the 60s, 70s, compilers were a new thing. People argued they were bad, that humans would always have to compile code manually. They were 10x slow and less efficient, like all the things, right? Okay, so what is the analogy to PCB then, right? So in our world, the PCB is like its actual design, the copper, the traces, the fiberglass, all that stuff. That's sort of like the machine instructions for the world of how to actually build and run that board, right? So we're taking a schematic. Judy Warner (18:20.562) Mm-hmm. Sergiy Nesterenko (18:22.35) And the primary job of the PCB is don't screw up the schematic, right? To faithfully recreate the intent of the schematic as the job of the PCB. But of course, in our world, to suggest that you do it automatically and don't review the layout is equally laughable as the, the converse statement in software engineering, right? So the question becomes why, right? Like why are, why is it a laughable idea that like you can click a button and a machine produces a layout that you just don't even review and send to. So in my opinion, and of course the question is why have auto routers not solved this, right? In my opinion, there's kind of three major shortcomings that have to be addressed for that to become possible. The first shortcoming is just that auto routers are focused on one specific problem, which is basically connect the dots. But as we all know, like your choice of stack up, your choice of placement, your choice of floor plan, your choice of materials, your choice of copper thickness, so on and so forth, like impact that problem. Judy Warner (19:08.253) Mm hmm. Sergiy Nesterenko (19:20.112) And so already you're doomed to fail because you're tackling only a subset, right? And you're by definition not tackling other parts of it and other parts of it that have a large influence on the routing. So that's problem number one. Problem number two is that auto routers in general don't complete the job, right? If you give an auto router a board with, I don't know, a hundred pins and it's sparse, sure, it'll hit a hundred percent and it'll be manufacturable. Judy Warner (19:25.818) Indeed. Sergiy Nesterenko (19:43.717) But if you give it something of any reasonable complexity, something that professional designers actually design, I mean, you're lucky if it gets to 70 or 80%. Right. And the thought might be, okay, well, it got to 70 or 80%. Like, isn't that 70 % of the work? Like, don't you just finish the other 20 % and you've saved half the time? Very much not. Right. And a part of that, and we'll get to the physics in a second, but a part of that is that the pins it leaves behind are by definition the hardest, right? It has blocked off access to those pins. Judy Warner (19:50.876) Mm-hmm. Judy Warner (20:10.107) Right. Sergiy Nesterenko (20:11.939) even as a human, can't get to them anymore. So you have to delete what it did, get to those pins, redo what you deleted. And from personal experience, that is such a painful, painful thing to go through that all of us curse them and laugh at them. Right. And the third thing, in my opinion, the most important is that the output that a auto router or a compiler has to produce in the space has to come with a guarantee of functionality, right? It has to actually work. Judy Warner (20:24.39) Totally. Sergiy Nesterenko (20:38.053) Right, if we had code compilers that gave us code that literally gives machine instructions that probably didn't work and had a bunch of bugs, and to go in and like edit the machine instructions to get it to work, I mean, we'd probably still be writing machine instructions from scratch, right? So they must understand physics, right? They must understand the signals that are traveling on the board. They must understand the requirements for those signals. They must understand the threshold at which they work and don't. And they must come with a report, a guarantee that, hey, Judy Warner (20:50.36) Indeed. Judy Warner (20:56.176) Hmm. Sergiy Nesterenko (21:07.149) I've checked all of these assumptions, all of these tolerances, all of these physics effects, and they are all below the threshold at which the board would no longer work. I think if we solve those three things, tackle the whole problem, drive to completion, solve all the physics, then, then we have something where you might consider taking your eyes off the layout, right? And truly, truly, truly bringing a compiler into this world, right? So that's what we're driving for. We're driving for, for that kind of future. Judy Warner (21:36.573) Well, it sounds like a dream and I think it's interesting that you're looking at the physics, right? And the best engineers I know, whether they're specialized in signal or power delivery or integrity or whatever, they at the fundamental core are looking at the physics. And when you were, when you are unpacking what you just said, my mind is like, I'm thinking through all the things, particularly when I was doing RF, people go, how come what I simulate doesn't match your board? And I'm like, my gosh, how long do you have? Like just the handoff from design. Okay, we got the schematic intent perfect. I'm like, yeah, all your traces were perfect and the physics were perfect, but they're not. We're running it through etchants and chemicals and there's fiberglass. so are you saying that you're looking at the All of those physics that I mean. Sergiy Nesterenko (22:38.075) So I'm saying we'll have to eventually, right? So like, I wanna be very clear about what quilters long-term mission is and how we approach it versus what we've already accomplished, right? So yeah, yeah, yeah. So long-term, yes, like we have to look at all of the physics of the problem. Judy Warner (22:42.554) Right. Judy Warner (22:47.3) Okay. Okay, I'd love to hear more about that. Sergiy Nesterenko (22:57.871) Right. And, and look like, you know, when I was working on, on Falcon nine, Falcon heavy dealing with how do we make sure that, NASA astronauts or, or critical, you know, military, payloads don't crash because of the radiation effect. Like we also have to simulate, couldn't just fly at a bunch and see what happens. Right. And so there's an approach, the approach that I take at the company and the approach that I think makes sense in a world like this is to approach truth from the side of conservatism. Judy Warner (23:15.8) Absolutely. Sergiy Nesterenko (23:25.603) Right. So you are not going to a priori go and model everything. I mean, least of all, like the details of etching and like the imperfection on the edge of the traces. mean, that's, that's hardcore. Right. but that is where you can characterize and, understand the variance that you get from that and the impact you get from that and be conservative, right. Which is exactly what designers do, right? This is exactly what engineers do. so as far as where we are and where we're going, right. Judy Warner (23:35.566) Right. Right. Judy Warner (23:46.66) Right. Yeah, you bet. Sergiy Nesterenko (23:52.912) So for us, the core approach is really, really critical, right? Like we are tackling the whole problem. We work on boards where we can hit a hundred percent and we tackle the physics, right? Those are really important things. But of course we can't eat the whole elephant at once, right? And for this to ever actually work, it's not enough to just assemble a bunch of smart people, put them in a room and task them with a problem. You have to simultaneously build a business case. So you have to find areas you can tackle, can solve for real practical businesses, use that to... get more funding to attract more smart people, to solve more problems, to then bring those to customers, to get more customers. And you have to do this flywheel, right? And in my opinion, this is actually the hardest part of building something like Quilter is like finding that tightrope to walk is like, where can you start? What can you solve first? Where can you make an impact? How are you clear with what you cannot do, right? That's also really important so that you earn trust and do that over time. So. For us, frankly, just because of my background, because of what I did at first in my career, where we ended up is with test boards. So at SpaceX, for example, we do component characterization. We take ICs and put them into a radiation beam on Earth and see what happens to them. How do we measure everything that's going into and out of them? And as we hit them with protons, do they latch up? Do they have a bit flip? Whatever else. And these are just like big boards that have grids of components. that then get a lot of instrumentation. And more often than not, the boards are not that complicated from a PCB design perspective. These are not 20 layer, UHDI, BGA, crazy material Rogers designs. These are four or six layer kind of designs, but they're time consuming. Because you have a lot of sensors, because you have a lot of instrumentation, you have a lot of power. These typically, two, three weeks to design or something like that. So the nice thing about that is that those boards are made in low volume. Judy Warner (25:29.573) Mm-hmm. Sergiy Nesterenko (25:47.899) So we're not dealing with like yield and problems with, you know, millions of units. they're typically, you know, four or six layer. So nothing crazy in terms of manufacturing. They're typically fairly sparse, which means you don't have to deal with like really high density problems just yet. And the physics is well constrained, right? There's like maybe a few impedance control signals, maybe some high current signals, stuff like that. But you're usually not dealing with like really intense length matching and know, multiple DDR five chips and whatever else. Right. So, okay. Let's automate that completely. And like, what does that mean? Right. And we've done that today and we have automotive and aerospace customers that utilize that. And so, okay, for that application, we've solved it, right? Like you can use the, you can use Quilter, you can generate these boards, you can test a lot more. And it's a huge unlock for those, for those teams. And the question becomes like, where do we go next and next and next and next? Right. And so my job really is to figure out what are those little, little hops that are reasonable with every step that push us to our ultimate goal, which is one day this is automated. Judy Warner (26:49.244) It's a lofty goal, but just listening to you, Sergi, think it resonates with me anyway, over my 30 years of, you know, the things that I've seen. makes a lot of sense, you know, and something I hear you talking about, the, well, I wanted to back up and just say there's nothing more common than really brilliant Berkeley, Stanford, MIT grads that find a tough physics problem to solve, but they don't ever think about the business case. So, you know, I really appreciate that what you've done is think about how the money flows so it can actually get done. You know, that's the goal. And so I... I hear the wisdom in looking at both the business case and doing it iteratively over time, which by the way is Elon Musk going, we're going to Mars. It's not going to be in a day. It's going to be iteratively over time and through simulation and gain knowledge and all of that. So kudos to you for that. Sergiy Nesterenko (27:58.267) Yeah. Judy Warner (28:13.42) you're using AI. So let's talk about that a little bit and maybe how it's useful now and it's not fully baked. Obviously it is also in its own rapid iteration. how are you able to use it today? Sort of how do you see it maybe coming to bear as you look ahead? even just a year or two, like how do you manage that Sergey? It seems like a lot and it's controversial in our little minnow mini ecosystem of board people. There's a lot of naysayers. So how do you manage that whole, the controversy where it really is keeping your feet on the ground where it really is and thinking about as a business owner and CEO where it's going to be, how do you manage all of that? Sergiy Nesterenko (29:07.974) Yeah, yeah, it's a great question. So let me first define what we mean by AI, because I think different people mean different things. And then we'll dive into the next couple of questions. So usually when I speak to someone, their first assumption is to make AI synonymous with like, chat GPT. That's like the most accessible AI that I think most people have spent some time with is you log into it, you have a chat box, you talk with it. And like naturally as an electrical engineer, what are you gonna do? You're gonna like ask it circuit questions, you're gonna ask it SIPI questions. Judy Warner (29:12.728) Okay, indeed. Judy Warner (29:24.294) Ha ha ha. Sergiy Nesterenko (29:37.127) You'll be like, draw me a layout, and then you'll have some chuckles and all those things. And that's all fun. Now, that is a subset of what has become known as artificial intelligence. And I personally don't actually like the phrase artificial intelligence. I think it oversells what's really happening. And it obfuscates, and it prevents first principles thinking, which I'm obviously very much against. And frankly, the term neural network falls into that same category for me. I think it makes it think that we've made an artificial and it's a real human. like, that's not a useful model for what's happening, right? So the way I reduce what a neural network is in my head for my ability to operate and evaluate is to treat it as an interpolative database, right? So what I mean by that is that surely everyone in school has fit aligned to some data, right? Like you have a Y axis, X axis, you have some points scattered, you find a regression, you know, get R squared as close to one as you can. Find MX plus B and then great. Now you can make some predictions about that data. Fundamentally, a neural network is nothing more than that. It's just far more than two axis, more than X and Y and far more variables than and B, right? Potentially many billions if not trillions. Now in chat GPT, it has been applied to language, right? So the way that it was initially trained was go scrape the world's internet of text, read a sentence, predict the next sentence. Judy Warner (30:38.204) Mm-hmm. Sergiy Nesterenko (31:01.458) That's fundamentally what it's doing. So no surprise, it's excellent at language, right? Like one of the ways that I use it personally is in communication, right? Like I have my own kind of internal writing about like how I think about culture and what it's going to do and how we're approaching it. And if I need to write a summary, I can take what I've written, ask it for a summary, it an excellent job at like mimicking my voice and highlighting the important pieces and all that stuff. That's great. But obviously like language is not... Judy Warner (31:10.118) Mm-hmm. Sergiy Nesterenko (31:29.946) the domain in which you should work when you're thinking about PCB layout. Right. PCB layout is a geometry and physics problem, not so much a language problem. We can obviously converse about it in language, but it's actually quite difficult, right? Like when you're a schematic designer and you go talk with your PCB layout person, do you hold your hands behind your back and explain what you need? Or do you point at the screen and say, that should be further, that should be closer, that should be more, you know, whatever. So for us, language models for PCB layouts are not used at all. Judy Warner (31:33.936) Mm-hmm. Sergiy Nesterenko (32:00.784) Like there is no chat GPT or Claude or whatever underneath Quilter. They would be useful and I recommend to designers to start looking at them for like, you know, reading data sheets, for writing firmware, for communicating with other engineers, frankly, for being persuasive and building your case within a company for whatever you want to accomplish. There's lots of different good applications for language models, but layout isn't one of them. Okay. So... Judy Warner (32:20.761) Mm-hmm. Sergiy Nesterenko (32:27.827) When we think about what neural networks can do and how do we use them, we really use them as what I'd mentioned earlier, which is an interpolative database. It is a way to store a bunch of results and predict what might happen on similar results. So one of the ways that we're not alone in doing this, there's a bunch of research looking at this, is to kind of shortcut physics simulations. What you can do is... Say you have an electromagnetic problem like an antenna radiation problem and it takes a Like HFS running an FEM for 24 hours to really characterize it if you generate a large data set of different antenna patterns and then you can train a net to kind of approximate what the Simulation would tell and then if you get an antenna that's like similar to the ones you've simulated But not quite you have some hope of that neural net predicting accurately what the antenna pattern will be or the gain will be or something like that But you can all of sudden do it in a second Judy Warner (33:21.5) Mm-hmm. Sergiy Nesterenko (33:24.625) rather than in a day. So in a design context, that means that you can vary the shape of the antenna millions of times, use this one second heuristic approximation to figure out what antenna shape would be the best. And then once you've got it locked down, run the old school HFS process to verify. And if it's wrong, train a better AI. If it's right, great. You've just designed an excellent antenna, right? So this is like one of the ways that we can utilize it. There's more like that, but that's like the tool that we lean on. And thinking about this as a CEO and a business owner, I'll steal what Sam Altman kind of shares on this thinking, I think he has a good framework for thinking about this. It's you should build your business with the assumption that AI is only going to get better from here, right? That the way it is today is obviously the floor, and it's obviously in a very fast trajectory. And we could debate how fast that trajectory really is, but it's only going to improve. And so... If you skate to where the puck is going, you can build your business around that assumption. And that's just a reality of what I think is going to happen. Judy Warner (34:33.082) I love that it helps me. I'm a visual thinker to skate to where the puck is going, which was right in line with my question. Like, how do you think about it now and think about it later? Cause really in the position you're in. so I wanted to back up to something you said. So is our, so is quilter basically training up models in the way that you talked about with the say that you and I issue is. training that up and then using the tools to verify. So in other words, when you go into HFSS, you're so much closer by doing that preemptively. Is that what I'm hearing you say? Sergiy Nesterenko (35:18.995) So that's a small part of what we do at this point. To be honest, the problem of just where do you place the components and how do you route them in at all reasonable ways, it's self really difficult, right? So there's a lot that we can do with models just in that domain before we can even get to the physics. Judy Warner (35:28.507) Yeah. Judy Warner (35:36.06) Okay, so that's what I wasn't clear on. want to make sure my audience was with me as to so, so where are you today? Today, how you're applying AI successfully to help get the layout done. Sergiy Nesterenko (35:52.328) Yeah, these kinds of techniques, you know, using neural nets, using the optimization algorithms that go along with neural nets, they can be applied to basically games, right? So one of the kind of historic classical cases where reinforcement learning really grew up was in Atari games, right? So all the classic like Pong and whatnot, right? And so, Judy Warner (36:08.879) Mm-hmm. Sergiy Nesterenko (36:22.962) What you could do is you can create a game that is effectively a PCB layout game. So, and in a sense you can view traditional CAD like this, right? Like you have some actions you can make, like move components around, draw traces, so on and so forth. You can define some objective function. I have to complete routing and pass DRC for example. and then as a human, you play that game. Right. And so that's basically what we do. Judy Warner (36:29.883) Mm-hmm. Judy Warner (36:47.067) 100%. Sergiy Nesterenko (36:49.844) Right? So we've invested in building basically our own CAD system under the hood because you need a CAD system that is kind of optimized for an agent rather than optimized for a human. For example, it doesn't need to have a screen, but it's nice if it has some other properties. And then you give it over to an algorithm that basically learns how to play that game and gets better and better at that game. That's the fundamental thing that we spend most of our time applying it to. And the work really is to improve the quality of the reward, right? Like, so what does it mean to win the game? Right. One thing is just to say you've completed routing past TRC. It's another thing to say you've completed routing past TRC and verify that all of your high current traces are, you know, going to be not overheating and then RF and then so on and so forth. So one aspect of what we do is, kind of improving the reward function. Another aspect is. Judy Warner (37:39.621) Mm hmm. Yeah. Sergiy Nesterenko (37:47.989) within the game itself, we can kind of help the AI out by making the game easier to play. Right. And so what I mean by this is like, you know, if the game requires you to make a million moves in perfect sequence to win it, that's a very difficult game. But if like the game is pretty easy to beat with, you know, even sometimes making bad moves, but it kind of helps you or corrects you. I mean, this is basically cheating in video games, but it's good for us. Judy Warner (38:02.907) Mm-hmm. Sergiy Nesterenko (38:16.18) then great, then it's easier for the agent to win that game. So we can kind of construct heuristics, shortcuts, aids, you know, kind of like PCB designers have to win the game, right? So for example, like in Altium, can, when you're placing a component, you can force Altium to make sure that two components don't collide as you place it. And that way you, by construction, avoid a DRC error with two components colliding. Like that sort of thing turns out to help quite a bit, not just for humans, but for agents. Judy Warner (38:35.685) Mm-hmm. Sergiy Nesterenko (38:44.948) So we spend a lot of our time doing that. Judy Warner (38:47.461) fascinating. And my mind goes to thinking, like you said, know, decades ago, everybody was freaking out about, no, you have to look at the machine code, you know, and, and I see you teaching, teaching the system to truly put in, you know, that learning. You know, I've heard other people say, which I, I, you know, just on mainstream is like. to your point about not liking the word artificial intelligence, it is misleading. And what I've heard people say is, it's actually not intelligent yet. You know, it does some parlor tricks with language, which is great. I'm like you, I use it all the time. And mostly the shortcut, right? It doesn't take away my work or my thinking, but it fast tracks the tasks that I need to get done. so I think that idea of training it to get intelligent, is something you're really leaning into and it makes a lot of sense. And I think it'll make sense to our, to our audience. Now, as we wrap up, I have a couple additional questions for you. When, when I named something I've been observing, like again, Sergey, started way back in the eighties. I'm dating myself. till now and I've seen so many changes in that time and so many times we hit this and we'll never do this and we break through, we'll never do this and we just keep breaking through. So this has formed my own perspective and it's like 100 % AI is gonna impact board design. Just because of what I've seen over time but I think one of the biggest obstacles Judy Warner (40:49.045) I've noticed over my career just talking and listening to people and working with engineers is we've complained for years for being in these silos because there's a lot to know just about signal integrity or power integrity or board design or circuit design or tests. know, there's a lot to know about each of these disciplines, but we've been kind of cut off to our fellow stakeholders either down or up. the supply chain, my observation, I'd be interested in your perspective is things are so fast and precise and high performance now that we have to get out of the silos or at least give digital visibility up and down the stakeholders or we're in big trouble. And it's why I named the podcast, the ecosystem, because I want to have helpful discussions across all of the stakeholders in electronic design and development. And they'll be like, I didn't know the test engineers needed that. So hopefully facilitate that. And I want to do that here too. So as you're building Quilter and you're solving this one problem, this one big problem, and you'll be solving it for some years to come, how do you think about framing the PCB design with informing it with that systems thinking. Sergiy Nesterenko (42:20.117) I love that. You might be the first person to have asked me that question. That's a great question. I thought a lot about this, not enough people ask it. So I think that that's really great. I think the cool thing about PCB design and PCB designers as people is you're actually at the center of so many different things at the company. For example, right? Like the number of people I've talked to that are like, mechanical team wants to move a connector. Judy Warner (42:41.595) You Sergiy Nesterenko (42:49.427) Yeah, what does that do to PCB design? And you've already done much of it and the redesign and now it's a debate or there's a procurement problem. You have a bomb change or there's a firmware request and now you're breaking out a different part of a microprocessor or a form factor change. I mean, there's just like a million different things. So if you actually map out like who are all of the people that could push or even a small request seemingly, and it ends up on the PCB designer's desk, it's dozens. Judy Warner (43:03.077) There it is. Sergiy Nesterenko (43:18.389) It's dozens. And this is a core problem is that like, man, accommodating those is really hard, right? You, you try really hard to collect all of them upfront. You design with all of them in mind. Inevitably somebody comes to you at the last minute, ask for five different changes. And a lot of times you have to say no. You don't want to say no, you don't want to hold back the company, but the reality is it's three more weeks of layout to like make some change and undo and everything. Right. So again, coming back to the analogy of software. In software, this is actually also obviously a problem, but it's also been really well solved. Right. So in software, you also get kind of hundreds of people potentially contributing small little things that compile into one system that becomes the browser or the video app or the operating system or whatever. And the beautiful thing about the way that works is that each of those people can propose that change almost individually and then verify everybody's input almost automatically. Right. So this process called continuous integration, continuous development. Judy Warner (44:01.2) Mm-hmm. Mm-hmm. Judy Warner (44:12.388) Hmm. Sergiy Nesterenko (44:14.963) where as a software engineer, if I want to change the color of a button on my website, I go up in the code, I run it locally, I change a couple of strings, I push it to our integration system on Git or GitLab or whatever GitHub. then like thousands of engineers will have written thousands of different tests to make sure that that change doesn't break anything. We have a review process and we merge. And like when you look at the best software companies, think Stripe gave this example where they push I don't want to misquote, but it was like dozens of major changes per day, or maybe hundreds even. And they serve like one in 10 people in the world with financial transactions and their uptime is 99.9999%. Like it's, I'll get the stats, but it's, it's bananas, right? Like to think that, yeah. Yeah. And so one of the exciting things about quilter is like, that world is uniquely unlocked with the existence of a compiler. Right. So in a world where Judy Warner (44:57.785) That is bananas. Judy Warner (45:10.704) Mm-hmm. Sergiy Nesterenko (45:12.625) Okay, like a schematic designer comes in and makes a little change and that audit layout is automated. Boom, you start a run that reruns the layout reruns all the physics checks reruns all the manufacturing checks gives you a report. And if I made that schematic change and all those checks passed, I can then go to my fellow designers and say, Hey, this looks like a good change. I need it. Work is done. You guys accept approve done. And all of a sudden you kind of decentralized that. So I think that tooling and automation not only kind of saves us time and you short its budgets, it's actually a communication system. It's a way to decentralize this decision making and allow multiple people to contribute completely independently and kind of achieve a much, faster pace. Now, hardware will always be a little bit different than software, of course, but there's a lot that it could improve on in this kind of world. And I think that that's a really exciting future. Judy Warner (46:03.791) I think that is probably the most exciting thing I've ever heard when it comes to hardware. Honestly, I do. Cause I think that's the right answer is how do we enable that? How do we unlock that? Because the way that the industry at large is trying to solve this problem, even at its best, Sergey is pretty janky and pretty, I don't know. It's, I don't know, putting lipstick on the pig a little, you know, I, it's like, I mean, there are digital bridges and APIs and things that we're doing and they are helpful and they're good, but it still really isn't solving for that fundamental problem. And it sounds like to me that you're going all the way to the bottom. and yeah. So anyways, very exciting. love the way you're thinking. Okay. Last question. I've kept you for a long time, but thank you so much. I've so much enjoyed our conversation. Looking into your crystal ball and seeing all that you see right now, which is fascinating and it's a lot. Where do you think, do you think we're going to get to that place where it gives a more fluid systems wide view? Where do you think we're going? Give us a short range, like a year, two years, and then give us Mars. Sergiy Nesterenko (47:37.885) Okay, yeah, I like that. Yeah, I mean, look, I think that fundamentally for some teams, this is a reality today. Right? Like if the kinds of test boards I was describing earlier, there are teams who using Quilter for which that's already the case, right? They focus on their schematic, they make a schematic change, problem. Click a button, layouts there, still some cleanup, but it's 30 minutes of cleanup versus three days of layout or 10 days of layout. for some teams, for limited teams for which the challenges are appropriately bounded in a way that we can already solve, that's already the reality. Now, granted, this is still within the world of electronics engineering, still in the world of a layout designer or a schematic designer or a firmware engineer. It's actually a fun use case. A lot of firmware engineers come to us and say, can I just like get a quick test board while I wait for the double E's to make the real one, right? That's already starting to happen. Yeah, yeah, yeah. The reality is... Judy Warner (48:28.897) I love that. That is so cool. Test boards are a big deal. Sergiy Nesterenko (48:34.88) They're a deal. And the reality is I found it in my time at SpaceX and in general, like there's 10 times as many people who know how to read and write schematics as those who know how to open up and drive a cat tool. And so all of those people can't make a board, but they know the electronics to make a board, right? And so it's a huge unlock and that's a lot of where we see our demand. But I think like, you dreaming big, right? Dreaming Mars 20 years. Judy Warner (48:52.453) Right. Sergiy Nesterenko (48:59.798) long-term future, obviously we'll improve, we'll be able to do this on really complex boards. And we're approaching some level of complexity, but we still have a ways to go before we get to server cards or whatever. But I think what will happen is you'll see this in other engineering tools. Already, obviously, we're seeing this in code. Code agents are probably one of the best applications of certainly large language models. And it's at the point where you can... send a Slack message to an AI agent and it'll make a merge request that you can review and accept. like for small to medium sized changes, it works. It actually works. Not big ones and it's limited and same thing as us, but it's there. I think the same thing is going to happen in mechanical. think the same thing is going to happen in thermal design and fluid design. And like the 20 year future is like, we as people are teaching these individual AI agents, new tricks, new manufacturing processes, new kind of concepts. And Judy Warner (49:35.013) Mm-hmm. Mm-hmm. Sergiy Nesterenko (49:56.523) baking them into the toolkit that these tools have. And then our representatives, like these agents, will then go and debate with each other how to make the product. So like, quilter will be arguing with the mechanical quilter and the software quilter and the fluid dynamics quilter about some trade space. they'll be making some decision about what the best product could look like and then presenting to us humans what they found. And so I think that that's... What I see as the potential, right? Where humans can do what we've historically been the best at, which is like true invention, right? Like true new ideas, true first principles thinking. And then anything we've learned in any manufacturing process or technique or approach we've learned gets automated and baked in and automatically assessed as a viable option for designing something across all of these different electrical mechanical systems agents. And then they debate and figure out here's the best product. Judy Warner (50:32.089) Mm-hmm. Judy Warner (50:51.823) Will AI in this world always be assistive? Like everybody is afraid they're going to lose their jobs and there'll be no more engineers. you know, what do you think there? I I'm with you in that, everything I'm seeing is absolutely engineering in the loop, both now and for a very long time. But tell me what you think. Sergiy Nesterenko (51:18.74) Yeah, I think the reality is I think it's a complicated question to be honest, right? Like I think history of automation shows that even when you do automate an entire job, lots more jobs pop up that are more valuable. But it also shows that design intent stays with the human, right? Like no matter the automation we've done, it's still the human coming up with what is the problem to solve? Who is the customer? What is the major approach of how to approach it? price point to quality ratio do we want to hit, right? I don't see that going away anytime soon. So, you know, in a hundred years, if we've achieved our mission and Quilter's compiling PCBs, maybe driving PCB layout and doing tracks is no longer something humans do. But does that mean that person's losing their job? I mean, probably not, right? They're probably going into something immediately adjacent, right? Which is designing the schematic, for example. And even if that gets abstracted to a sense, you're still designing it at some level, even if it's an architectural level, if not a detail level. Judy Warner (52:18.072) I'm old enough to say this is true because I've seen, yeah, if you live long enough, you see the alchemy that things will die off, but then five things will sprout up in its place. And you can see it very obviously in the music industry in that, you know, it was LPs and then it was cassette tapes and then it was CDs and da, da, da. And everybody said, you know, music's gonna die and da da, and look at it now, there's more people in the music industry and there's more concerts and now we're going back to LPs because they're vintage. And that can be extrapolated across many, many industries that we've seen. So I share your optimism from my limited perspective. Well, Sergi, this has been... I could talk to you for an hour and I'm sure our listeners is so enjoy listening to you sort of riff about this. where can people learn? I'll put links to Quilter in the show notes. you, is there anything coming up where our audience can connect with you? Maybe, I don't know, do you have demos of your software? Are you going to be on site anywhere? Tell us where the audience can connect with you. Sergiy Nesterenko (53:41.195) Yeah, easiest way is just to go visit our website, Quilter.ai. From there, you can try it out, right, for free, for non-commercial and for academic use cases. You could just use it, see what happens, right? See if it fits your use case. In that workflow also, if you wanna talk to us, there's a way to schedule a demo with us. That's no problem. Or you can always just find me on LinkedIn and reach out and just speak with me directly. I always welcome. Judy Warner (54:03.514) Great. Okay, I'll make sure I put all those links down in the description. Sergi, thank you so much. Thank you. think you're contributing in very significant ways in the industry. And again, I'll keep my eyes on you and the Quilter team. You're doing fantastic work. Thank you for what you're doing and we wish you continued success. Sergiy Nesterenko (54:23.905) Thank you, Judy. Appreciate it. Judy Warner (54:25.819) For our audience, thanks so much for joining today. I'm positive you enjoyed this conversation with Sergey Nesterenko. Please go check out the description and follow up with those links. We appreciate you so much. We'll see you next week. Until then, remember to always stay connected to the ecosystem. Baboomy.