How Diode Is 10x-ing Hardware Design

Key Takeaways

[Music] I'm excited today to welcome the founders of Dio, Davidid and Lenny. They just closed their series A for 11.4 million led by A16Z. And you guys just went through the batch just a year ago in the summer 24 batch. So pretty quick growth. What are you guys building? >> Diet is using AI to automate circuit board design. You can think about us as AI enabled design shop for printed circuit boards. >> Who are the customers that are using you? So, it's been a pretty rapid growth within just one year. >> Yeah, we we've been very lucky. We work with Fortune 100 companies as well as larger like uh startup companies like physical intelligence which is robotic foundation model company and Seronic which builds autonomous boats uh as well as like smaller startups that need board design. >> Take us to the beginning. So just a year ago both of you were still in your previous jobs when you had applied to YC and in fact you were still in your jobs few days before the batch started. Tell us about that journey. What what were you doing? How you decided to work together? >> Yeah our origin story actually goes quite a bit before that. Um so back during COVID when I was uh still in college I was a software engineer at this company that was also building hardware and I realized like kind of the interesting problems at this company were happening on the hardware side and so I asked to join their hardware team and had the good fortune of getting assigned a day as my intern mentor uh who graciously took me under his wing and showed me the ropes of what it's like to be building hardware um and I kind of got exposed to to this world of building physical things. Um and thought it was super exciting and super fun. So fast forward to you know the previous job where we were working together. Um, we were also working on a hardware product. I was, you know, more on the software side. Davity, you know, a bit more in touch with the electrical side. And, and we found that despite working with like, you know, the some of the best engineers I've worked with in my career. Um, you know, we'd find that, you know, circuit boards would come back with like silly mistakes in them because the tooling wasn't there to help you out. And as a software engineer, I'm used to being able to make mistakes, find them in 30 seconds, and fix them. And I realized very quickly that in electrical engineering, that's basically not how things work. And that was kind of the origin of how we decided to kick things off and see if we could do something to improve this state of the art. >> We had some experience designing u custom silicon and we saw like incredibly talented engineers basically oneshot these very complicated uh like TSMC designs and we just wanted the same for printed circuit boards. We we just found that the tooling was not um up to par. I was personally like very enthusiastic about how large language models uh can start to parse the documentation and c like they were able to catch those silly mistakes but you needed to give them like a representation on like to understand what the electrical design would look like and so we effectively applied to YC um with this idea of will give us your electrical design and we'll spot the mistakes. You spoke to over hundred people in the first couple weeks of the batch, have all these user conversations and you found out people didn't really want that. >> That's right. It was actually quite um quite interesting and very humbling. Like the original problem statement was I like am able to design printed circuit boards, but I want the software to tell me if I'm making some mistakes the same way that a compiler will tell me if I'm making mistakes while writing software. But when we went out and like pitched this idea and we we really built a lot of like infrastructure and like interesting stuff like we had this generative pipeline that would like create boards and then inject mistakes into them we could use to like reinforcement train the algorithm that would catch them. People basically told us we don't make mistakes like we don't need this. I'm sorry. Whether that's true or not it doesn't matter but it really speaks to the fact that like talking to customers is really really important. And so like at some point we had to recognize that the the pain point was not finding the mistakes within the uh existing design but it was actually like generating that design itself. >> And there was some interesting happen stance from you guys just being in the YC batch that you got more lucky to find the right problems because you were surrounded by all these ambitious fellow YC batchmates that you then stumbled into the actual idea for diode. tell us about that journey. >> Um, that's right. After like uh an office hours where we actually told you like, hey, like I don't think that the verification stuff is working out. We will we'll need to go back to the drawing boards. We were very somberly like having lunch with a really good friend of ours uh who who happened to be selling like uh edge models to robotics companies. And like he basically told us look like I don't know if this can be useful to you but we consistently hear that all of our customers seem to want a custom Jetson or uh like development board and like the Jetson or is a edge GPU made by Nvidia like very common for like developing robotics application and we realized this would be like very easy for us to do and very easy for us to generate and from that like kernel of information which I don't think we would have really stumbled upon so quickly if we hadn't been surrounded by other YC batches like batchmates. We went we went out and like closed our first deal like right away. That was like a really interesting moment from an other YC company. >> You got an instant yes on your first user conversation whereas the previous time you talked to over 100 people and they were all basic lukewarm pretty much nos. Right. >> That's right. That was a really good validation. The interesting part is that people didn't want a component of the solution. They wanted the solution. they didn't want just the tool that makes their job a bit easier. Like the biggest pain point was that there's not that many people that can actually design circuit boards um in the US. Uh and so for for a lot of companies, it's hard to like find the right talent and like bring it in house in the early stages of like prototyping and even for larger companies, it's very often that they outsource this type of projects if you don't have bandwidth to do all the things that you want to do. And so it turns out that uh diode can really help with that. like generating boards is where the pain point is at. >> Which is super interesting because in the process of building the software to verify circuit boards, you had built one part of the type pipeline that would generate boards to really go through the iterations of them and that's what people wanted. >> That's right. >> It was a bit counterintuitive as well because you both I remember we had this office hours and you were confused. It's like how do we even build this into a venture scale business, right? because you are effectively selling at that point services to design boards. So how do you guys square that? How how did it come about? >> I think that um the realization came from the fact that for the first time in a very long time we can design software that makes us incredibly productive at like solving the problem end to end. And the 8020 problem like a lot of products in the AI world kind of fail because you will never be 100% correct in every single thing that you do. Uh but when you are like an expert in a specific domain and you can solve the 8020 problem um it now becomes possible for you to actually like package this solution and sell it as effectively like services as a product. It's working in a way that um we hadn't anticipated before. uh but we really take ownership on our product and the output of our work and this is what our clients demand in terms of solution like we use AI to automate the internal working of diode but then this is not something that we just put out into the world without verifying that the output is actually correct uh we own the verification pipeline and we make sure that anything that we deliver to a client will actually work that we have like acceptance criteria with them uh it very much resembles like traditional development it's just faster um and uh more efficient. >> Yeah, I think the core insight here is if you could reframe the PCB design problem as a software problem. You can take all the great learnings that have happened in the world of software both in terms of verification and compilers that will tell you if you're making a mistake as well as applying all the great research and and work that's going into automating software engineering and uh make that accessible to the world of PCB design as well. And you're basically able to benefit from from that with the work that we do for our customers. So I think the really cool thing is you basically have software margins for this PCB design services because of AI, right? And this is only possible now. Tell us about what would happen using the largest models and latest one. >> Yeah, I think the interesting thing is like if you look at those models today, they know everything they need to be electrical engineers. They have the innate knowledge of uh how circuit boards work, how they need to get wired up, the rules that electrical engineers spend years learning as well. They just like don't really have the mechanism to do the work that an electrical engineer does cuz traditional tooling is very visual and graphical. Hasn't really changed much since maybe the 80s or so. And so all you really need to do is basically just convert it into a form that's similar to what they're used to doing, which is writing code. And then all of that knowledge that they have can start being expressed into the boards that they design as well. And so you can kind of unlock this like latent capability that these models already have. >> Um, another interesting learning is that one of the reasons why this hasn't happened yet is that traditionally like incredibly smart electrical engineers have been trained in this very visual way. And so there are some benefits even for humans to like reason about things in code. That's why like um RTL or the regist transfer language for custom silicon is written in code. But the the the the workforce we have currently is very like traditionally trained in visual processing. And so because the models are really the ones writing this code, we can now let them write it and just like export a visual representation to our users, which is what they're used to. So we don't need to actually teach humans that have incredible domain knowledge about code. We can just like generate it in the back end as a implementation detail. I personally happen to think that like reading that code is very enjoyable. Uh but we we don't want to force it on anybody and so the output that we give is very traditional but it just allows the model to have an intermediate representation that they can modify. >> I think one interesting stat you shared is just the two of you you've been able to decide design over 100 boards in the last couple of months which is crazy. A regular electrical engineer would only be able to output a couple. I think it depends on the complexity of the boards. Um, but there are incredibly talented engineers that are also fast. Um, but we have to manage a lot of different clients. Uh, and it would be impossible to do uh without the like software tools that Lenny and his team build inside diode. So like the way that we divide the work inside diode is that I had the electrical engineering team, Lenny has the software team. And so we are effectively their first client. And in the early stages, especially during the batch, it was just the two of us. And so he would like craft these incredibly high-end software tools that I would use to speed up myself. Um, that was incredibly valuable. I I've always personally like really wanted to have these tools. Uh, there are some open source tools in the world of electronics like Kik, which are incredibly like big fans of um, but he really built a fantastic layer on top that allows us to go even faster than that. So you've done very interesting technical choices to be able to serve in these very old school industries selling to aerospace medical device. That's not a easy feat because a lot of them tend to be very conservative with adopting new tech. What are the technical choices you've done that you actually close this large Fortune 100 and all these cool tech companies in hardware? Yeah. So I think there's a lot of really interesting technical questions and problems we have to solve. Um the first of them is even just designing the language that we use to represent the schematics. We have this kind of novel challenge of designing a language that works really well for humans and also works really well for LLMs because we want LM to be able to generate this code. And LM have, you know, millions and millions of lines of code in their training data. And so if you design a language that's very familiar to them, they'll do a much better job at not making syntax errors and kind of writing code in a way that's natural to them. But at the same time it has to be something that's intuitive and readable for the humans like on David's team to go and review the code, write some of their own as well. Um, so I think that was kind of an interesting set of challenges we had to make. And then once we design the language, we have to design the infrastructure in a way that's easy to compartmentalize and deploy in settings where there's really rigorous security requirements. For example, working with aerospace companies where it needs to run on an air gap system, for example. Um and so we have this kind of really tight rust compiler which is uh implements the core logic of our um implementation but we also then have you know bindings to was so we can run this in the browser and have nice visualizations and compile schematics in real time locally um for our users to review. Um we can automatically generate configuration blocks for configurable modules where users can go on the website change the configuration of their buck converter or what have you and see the schematic get recompiled in their browser in real time. And this is all kind of enabled by this kind of uh really tight isolated like offline core component that uh powers the the infrastructure we've built. >> I think this is one of those companies that only you two could have built because it's a unique way of uh approaching hardware design for circuit boards as a software problem with with Lenny right and the experience on hardware. So what are you excited about the future? I I personally love the fact that I see the world around me changing incredibly fast and I see like AI really 10xing our software development speed. But the the thing that excites me the most is taking this incredibly powerful um like set of forces and actually applying them to uh like the physical world like designing physical object in physical space. I think this this is really the next frontier. This actually has been fantastic for recruiting. There's a lot of really smart software engineers that are looking for the next like Everest to like uh climb and and we really think hardware is the next frontier of like really hard challenges. Hardware and like generating these boards and putting them in the real world and making sure they work reliably and using AI to solve these problems is really what I think will like move civilization forward. >> What are the kinds of people that you want to hire? Well, I mean, I think on the software side, um, you know, we want people who are curious and excited to be exploring like entirely new frontiers with open-ended questions. Um, like I said, we have a lot of questions where we know what the problem we solve is, but the solution can look one of 10 different ways. And so, we need people who are excited to kind of go and and try experimenting and prototyping and answering these kind of open questions. Um, so kind of fun researchy type problems, but at the same time, we need to ship products, we ship boards, and so it's like very practical and applied. And I think the the other kinds of engineers who we really love working with are people who are excited to build cool experiences for our customers and users to interact with the boards we design, right? So, we have to be able to hand over the designs to our customers and they need to be able to inspect them, review them, give feedback. Um, and so building these kinds of like great interfaces, I think, is um also something we're really excited to press the accelerator on. >> Well, congrats against on your series A. >> Thank you. Thanks, Diana. [Music]