Latent Space: The AI Engineer Podcast
Original podcast
Watch the full show on YouTube: https://www.youtube.com/watch?v=_OeDFb3oIAY
9,451 words · 2 speakers
All right, we are in the remote studio with me, T S one there, CEO, co-founder of Fuxci, welcome.
Thank you, great to be on the show.
So it says here on the website, you are building the AI hardware engineer. What's your journey to building the AI hardware engineer? That's an unusual concept.
That's a great question. So I look, I've always been a builder from childhood on, you know, Legos, wooden blocks, whatever. I went to school for half a software engineering back in Germany, worked at Siemens for many years afterwards, started like two hardware software companies out of that. And then 15 years ago, moved to Silicon Valley. I was at Meta for a while, both on hardware software as a product manager. And yeah, you know, I left Meta in early 2019 and you know, I've also been going to Burningman for a long time and I decided, look, I'm going to check the summer off and going to work on Burningman projects, right? The whole summer long. And that's kind of how I got back into hardware after like lots of years being mostly in software. And I had this realization that the tooling to make hardware had not improved in my lifetime.
And that was so odd because the tooling to make software is like unrecognizable if you compare it to like mid 90s or something like that, right? It's like bonkers what we can do now. And I was like, isn't that odd, right? Because the supply chain for hardware has changed to such that I can as an individual now get pretty much anything manufactured, right? Why is the tooling not caught up with that capability, right? And I think like 3D printing was like an early trend here in the direction to like, oh, you don't need injection molding, you can just 3D put this at home with like a thousand dollars 3D printer. But that thing has been like slowly, gradually taking off. But in electronics specifically, nothing really has happened and that was kind of weird, right? And so I kept complaining about that to friends all summer long. And eventually a good friend of mine said, Matthias, look, you've been complaining about
this all summer, you either going to do something about it or stop talking about it.
I could do something about this. And so then we started Flux in October 2019. And I think the key insight here was the coming from Meta. Like Meta, right? We were like far ahead, using deep learning, machine learning to power like anything. You can imagine, right? And I had seen the power of that technology, right? And it was so obvious that there would be tons of use cases here to automate electronics design, PCB design, you know, with machine learning. And I wasn't even thinking of L L M's at that point. And so we set out to build like a smart design tools, what we called it back then. I thought about calling AI, but I didn't want to sound like a lunatic. And so, but then, yeah, when you look, you fast forward, L L M's happened in 22, right? And then the whole thing really took off then, right? And it's just, it was actually pretty good because, you know,
if you're building like an AI code tool, like a dev and a cursor or so, right? You can fork VS code, right? You can slap an L L M in and like, there's a lot of like infrastructure that you don't have to build, right? If like Cursor or Devon would have to build their own VS code, that'll cost a hundred million dollars, you know? Call me back in five years. And so now space, right? The opportunity didn't exist to just fork something like that because there was no open source design tool for electronics, like not something good enough, at least not in a browser, not written, you know, with modern architecture. And so we spent the first couple of years building it from scratch into the browser, kind of like a similar journey that like Sigmar did, right?
And made that with a collaborative and built that with in mind to use machine learning models, right? Built the data structures, right? Feedback loops, we always thought of the tool itself as the reinforcement learning environment, right? And we were just building the interface for users to provide us feedback and correct the agents. And then L L M came, so long ago, then, by whom, right? And that just turbocharged the whole thing. And so, yeah, the product today, right? You can probably come in, design things manually, or you can have the agent design it for you, or you can collaborate. And I think this is actually similar like in software engineering, most of the time we're doing is collaboration with the agent, right? The agent does research for you or figures out some stuff, right? Or runs some reviews and you can go back and forth until you have your, yeah, your manufacturable design out there.
Yeah, amazing. I think I got in touch with you because of Manny, which is he, which, uh, who is one of your angels, I think. And so like, what is the sort of really inflection points for you? I think you've been going sort of pre L L M's for a while. Did you just plug in G P T three and like everything took off? Or was there a more complicated engineering story behind that?
Yeah, it's a good question. So I think I'm gonna claim here, I think we were the first engineering tool or design tool that had an A I chat in it. We shipped that, I think, a month or two months before G P T four became publicly available. And the the realization we had is like that there's so much of the job in electronics design is information gathering. But you have all these semiconductors, the S D, the sheets for them, which one are we gonna use? What are the exact specs, how do we configure that thing? And that like an agent here with rack, right? Would like supercharge the process and you would not have to Google anymore. You can just ask the agent, right? And that was like the first product we shipped. And that was, yeah, I think that was like February twenty two, if I'm getting that right, or twenty three?
Something like that, right? And and that was like magical, right? At the time, what didn't work so good at the time is to want then, of course, the agent to do things for you, to take actions. And if you remember, like tool calling wasn't working that great in the early days. I mean, even sort of a chat activity that tried that early on and it would consistently pick the wrong tool. And that wasn't really working. And I would say like that tool calling, I think, really only became reliable about a year ago, like last October, November. When the the models that shipped were suddenly really good at calling tools. And I think that then changed everything, right? And you see that and then all these A I calling agents took off, you know, like massively.
Because suddenly you could use Grab, right? As a tool call effectively, right? That was like a big unlock. Because before, everybody was just like letting the L M edit the the freaking file, which was like very error prone and slow. And then for us too, right? We were then able to hook in all these tools, like suddenly you can have like a sub agent that's hidden behind the tool to search the library, right? Or or figure out pricing and and and lead times for semiconductors, right? Suddenly you could have tools to essentially gather the context of the current project to like solve the current user request, right? And you could have suddenly tools to add components to your project, remove them, wire them, run reviews, right? And all that kind of stuff. And so that was a big game changer. And I'd say that in the same time, what we saw that there was a big shift in what users
expected from an AI tool around January this year, I would say, well, like suddenly, like before users came to us, be like, can you support this manufacturing process? Can you support like laser vias and stuff like that? And then there was a sudden shift where suddenly everybody asked for, hey, I want to come with a product brief and just get a fully designed board out of this. I don't want to do so much manual stuff here, right? And that just overnight, and I think what is my hypothesis, what happened, I think in general in the world is that people changed what they expected from software, from using all these other tools like perplexity, chat GPT and so on, right? Cursor, it's like an overnight shift where like, no, I'm going to adopt a new piece
of software to do something for me, it needs to take in a prompt with the problem and come with the solution back to me. I'm not interested in anything else. And so that then led us to really like heavily invest in the agenda capability and flux, right? And going from like a, hey, look, I can answer questions and perform single tasks to like, no, I can take a product brief, break that down into an actual execution plan, execute that plan, adapt as I go, and then come back to the user with the fully designed thing. And then we finally shipped that thing in early October, so like a month ago.
Amazing. Maybe we should show a little bit of the products so that people can get a little bit of hands on. For me, when I log in, I see brainstorm new project of AI, ESP32 robot controller, Raspberry Pi Pico and Tesla Vigo display system.
Yeah, you'll see here, when you see my screen here, like you see all these like test products and running here to debug things. But um, you know, we have, we have some good examples here too. So maybe to start with, we've built a cat tool, like an electronic cat tool that's fully in the browser. It's like real time collaborative as the Google Doc or Figma is, right? We have like versioning in here, collaborative features, team functionality, all like good stuff, right? All these like basics. And then we have these agents in here, right? And so the information architecture we have is very similar to GitHub, right? We can fork things, you can like create branches, right? Collaborate here, merge back in. I'm going to just open here like a sample project.
Give this a second to load. And so then we have the full cat tool here in the browser. So you have your schematic here, you know, for those who are familiar here with electronics. And then you also have the layout editor here, the mechanical editor. And then you can go all the way here to, you know, visualize the board's going to look like. There you go. Alright, so and you can send this off from here to a manufacturer anywhere in the world. Typically China and they'll, you know, in a week, 10 days, they'll send this back to you fully assembled.
Is that, is there, is that something I can do on your platform or have to export a file and send it by normal means?
That's a great question. See, yeah, our goal is we want to seamlessly integrate manufacturers ultimately, but we haven't done this yet. You know, there is this like, there's a couple of different standardized file formats that you can give a manufacturer anywhere in the world. And they can manufacture the board for, for you from that. And you can just export them from Flux. The most popular one is called Gerber. I mean, the other thing we have here, I mean, just we have some windows so I can see it. You, we have, you know, the full supply chain integrated. So we can look here at what this board is going to cost. For this component, you know?
Wow.
What's it going to cost? Is it in stock? What are the lead times?
How do you know, how do you know what the costs are?
So, you know, we've integrated with all the semiconductor distributors.
Okay.
And we can look up what's in stock there, what will be the lead times. You can see here, error here in this case, right? They have zero in stock of this component. Digikey has 234, but the price here at unit count of one, right? And all this good stuff. And the cool thing is, right?
Matter the, the volume, because usually, you know?
So at, at volume one, it's $2.64, at 10, it's $1.66, you know, like 48. You know, it's a $1.25 and so on. Yeah, you have the, the price breaks in here.
And so that's great, but because the, the agent can carry this too. And then component being on stock is absolutely critical if you want to design and build something. If you think about legacy tools, you don't have these integrations, you don't have AI for that. But, but it's easier as a result to design something and when you have a design with 500 components. And one of them is not on stock or the lead time of 200 days. Yeah, then the whole project has a lead time of 200 days. Alright, and so that's a good problem and that's something that the agency also help in the tour. So for example, if we ask, let's go back to schematic. This is IC one. So let's, we can ask you, the agent is one of the simpler tasks, right?
What would be a cheaper replacement for as there?
Right? And so the agent can go now, look into the marketplace at all the distributors, right? Find suitable alternatives, look at within pricing, right? And come back to us. This is going to take here thirty, sixty seconds, hopefully. And it's like, this would have taken us hours to do manually, right?
Yeah. Are you using MCP under the hood by any chance or?
We don't use MCP, no. We, this is just all regular function calling, but we own the whole pipeline here. I feel like MCP makes more sense if you want to interface with third-party services, right?
Yeah.
But no, it's all function calling, you know, we use langchain, langraphy under the hood. So of course, answer with a counter question here. Let's do, yeah, we want a drop-in replacement.
And was that a, by the way, can you zoom in on the, on this left, left side? Because it's a bit hard to see, it's very small.
Was that a decision to do a counter question? Is that part of the process or the model just give up with it?
That's a great question. No, we go back and forth of that for like certain kind of requests. We want the model to default to action. So for example, right, if you're giving it an actual plan to execute, we're really trying to like, just jump into action, start working. If the user wants changes, they can just ask for that, right? But for other things like this one here, right? What it asks here is whether we want like a drop-in PIN compatible substitute. It, because that would mean that we don't have to redesign this whole thing. Right? And so that's like a critical question, you know, for our search.
But yeah, it's definitely something that's like, I would call this like personality design here, the personality UX of the agent is something we, we talk a lot about internally and innovate a lot about, right? So then you can see what it does here. We have our own library for all these parts. A million, that's all user generated. You know, it does a tool call now, but runs a hybrid search against that. And in some cases, so even like actually kicking in agent, like a plan execute agent here to be like, okay, look, here's this part I'm looking for, we're looking for alternatives to this, that are pin compatible. Let's brainstorm the different vendors that could be offering this. Let's brainstorm the different search queries, right? And then we'll perform a bunch of searches, look at those, evaluate those, and eventually come back to us here with the recommendation.
That's a very, very wide search space.
Yeah, yeah, it is. And the other thing it does is like, you know, sometimes the part might not be in our library. And so then it will do a web search, right? To find a suitable, you know, alternative that is in our library.
Is that again, do you use any third party tools that you made yourself like Exa is? It's your partner here in SF.
It's all our own stuff.
Web search, wow.
We use for the, for the vector search, we use Quadrant. Semantic search is Agolia, right? And then we just fire them off in parallel and see what they come back with. We're going to refuse the results then and re rank them.
Like I said earlier, all the agent stuff, that's all Langraph, Langchain, trying to single any other notable third party stuff we use. I mean, you know, OCR, we also don't do ourselves, we use Azure for that. So yeah, it came back now, you know. Another cool thing is like, you know, we have. Yeah, let's look, let's look, right? So look for pin compatible alternatives. And so yeah, pin compatible, you're only going to find other ESP32s, right? But yeah, offered to swap those, right?
And this, yeah, this is the common those. So let's ask you what the actual price for this is here. Um, I was like, oh yeah, the knowledge suggestions here. But so another thing with built yourself is like, you know, and user driven knowledge base. Where like as the agent learns things and it's what I learned here, right? It's like, right? What I learned here is like that, we're replacing ESP32, you know, we want a drop-in pin compatible thing, right? And we can now accept that, we can add it that, we can reject that. And what's really great about this is that it is enabled personalization on a user level, on a project level or organization level. And I think that's like critical now to AI systems and like, you know, I have a, if you ever use Tesla self-driving, right?
It's magical, but you have a construction zone for two months in front of your house. It'll just not pick up, right? Which lane to jump into and stuff like that, you know, because it doesn't have a personal, a real time personalization layer.
It doesn't have memory.
Right? And that's a huge issue for these, for these systems, I think. And so, yeah, we brought our own thing here and what's really cool is, like, the graph, the ability to centralize for your team to manage this knowledge, right? The agent creates the suggestions themselves and they're really good at picking this up, right?
So yeah, it's a game changer, but we also found that we have then system level knowledge here that we train the system to do. And that's really also how, you know, we use mostly the off-the-shelf frontier LMS here. And the way to like get them to behave the way you want them to behave is just with context engineering and injecting the right knowledge for the right kind of task they're doing. And that's incredibly effective, right? That reduces hallucinations, that makes the model answer in a specific way over and over again, not just one in 10 times.
Okay, that's amazing. And maybe I could ask for one more, um, sort of from scratch.
Yeah, yeah, let's do that.
Just, just because I'm obviously a bit newer to this, but I've, I've actually tried to do this. I've actually bought ESP32s and failed to do this. So my project is to make my own smart Alexa, right?
It needs a, it needs a, you know, beam forming microphones. It needs like on-device uh wake word um you know inference. And then it needs to be able to call um some API externally. And then it needs a screen.
Um yeah.
Is is that all within the requirements? I I don't know about the screen part.
Yeah, I know that that's that's that's actually great. So why don't we use that, right? So I want to build a NASA type device.
So like, let's say, uh speaker microphone screen. Uh and then some wake word inference. I don't know what else I need. Battery is nice to have. So the power draw is like a really interesting requirement.
Powered. And the battery should last a week.
Probably wifi enabled, right?
Oh, yeah.
Yeah, that's enough. You know, we don't, we don't need that too, too many requirements. Obviously, we could add things, but I just want to see like for the audience. Probably something they've thought about wanting. And they don't have. How quickly can I make it a reality, right? I think that's the, that's the real problem with a lot of software engineers who want to start with hardware that they can't really. They don't know what they don't know.
Yeah, absolutely.
Exactly what I ran into. I bought the uh ESP thirty-two with like some DevKit things and it's a bit slow and it's and then in the firmware that was preloaded was like really bad. And then I was just stuck.
Yeah, it's it's a little bit of a steep learning curve. I would say that like, you know, AI makes that a lot simpler today, especially like, you know, the debugging process.
Right?
As a novice, for example, there's like a thing that's like decoupling capacitors that you have to like set for like something ESP32, right?
Otherwise, you know, it won't work or it won't boot. And this is kind of stuff, you know, that's like, nobody points you at until you run into that issue. And the AI agent's actually really great at being like, hey, look, I built this thing. It's not doing the thing. Why is it not doing the thing? And the agent will be great, actually, at helping the triage, you know, and finding these issues and teach you.
It came back here with a plan. So let's look at this, right? Also, this is a battery powered Wi-Fi voice assistant, mic speaker and screen targeting one battery, one battery life. How to find a system?
And a low acceleration load switch to hit. Yeah. Oh, yeah, we didn't define how we're going to charge this thing. USB-C charging, fantastic, right?
Before, the wake model, always listening, voice work mode, a button to, yeah. So that's a good question, right? What wake model do you want? Always listening or a button?
Yeah, always this thing.
Yeah, right? That's what I'm thinking too. And then, yeah, the screen size.
Yeah, OLED, yeah, OLED.
Like a small OLED, it's great. And then cell types.
I don't know.
Yeah, yeah, this is great. And we could ask this now, but I'm just gonna go with here. The, it is a typical lithium ion battery, the 1865.
Let's go.
It would be, you know, it'd be interesting is like if there was like sort of little help buttons on each term in case you needed to like do a side chat.
Yeah, for sure. You know, we have some of that, but you're totally right. It's like this one would have been a great candidate for it, you know?
Yeah, exactly, because you knew that. For your previous context, but I did not know that I would have to search. But if there was like a little I icon, I could just click on it and, um, you know, then it would just tell me.
It's really interesting. You know, I was just, just last week, I was sitting here with Devon thinking through, or maybe we've kind of like a right click context menu here, right? Now it's gonna go do Google search here in Chrome. I'm sure you can see that on the screen share, but, um, to just like, you know, ask a follow up question based on this term, right? So I was just thinking in a similar vein, but I like this use case you have. You're just like, yeah, there could also just be an I button, you know, and that just triggers like an, uh, a pop up or so, you know, within
AI summary or so what this thing is. I love that.
Well, in, uh, Windsurf, we actually had this like, uh, thing where you can click on any symbol and it generates a deep wiki for you. Deep wiki is just kind of like a Wikipedia entry on it.
Yeah. I just reinstalled Windsurf last week for that. Also like the, what are they called? The code maps that just.
Code maps.
That was the, I worked on that launch.
Awesome. Yeah, I know. That's phenomenal. Now we use Devon here religiously and I think one of the most underrated features of Devon is the actual ask feature.
I use that even for my handwritten PR. So I'm like, no, let's just talk about this a little bit first here, you know?
Um, cause I think that's also a thing here, but I was gonna say is I think a lot of success, it depends how well you flesh out what you wanna do, right? Like if you just use this and ask it to like start executing a plan, it's, it's gonna be high variability and what you're gonna get back, right? But the more we go for here in specify, the better it's gonna be, right? And you see that here, right? There's tons of more questions, right? OLED interface, the so that, right? Speaker size, right? We're gonna ignore this now, just tell it to just get going. We can actually look at the plan, the execution plans created here. I mean, here, this size is.
That's why here's the goal of the project, implementation plan, you know, gotta do a schematic, gotta add all the components, gotta connect all the components. Then valid the power architecture and so on. So we're just gonna tell it, let's go.
Let's go. So first of all, before you, before you, this call, I actually didn't even know you used Devon, but yes, I also work on Devon. So it's kind of cool to see this.
We love Devon here.
Yeah, I mean, a lot here is inspired by Devin, like the knowledge-based thing, right? It's inspired by Devin, right?
Yeah, yeah, yeah. It's a very similar idea. So I was wondering, you know, if you had any references there. And, and you know, something that's interesting is the team has been working on computer use inside of Devin. And so you have, what you have basically is like, kind of like a, you know, cursor or Devin for PCB and like hardware design, right? And it's a very complex, dense user interface. And so typically, you know, Devin would not be able to like see this and navigate it. But recently they added computer use into Devin where it's got pixel precision on pressing buttons and typing in into the text field. So I'm pretty excited.
Actually this morning, I've yet to try it, but you told me right. So I'd say the previous version did pretty well navigating our interface where it struggled are these canvas interactions.
Right? Especially, I don't think it is able to do dragging gestures, like mouse dragging gestures. This is an issue for like a canvas-based thing.
Yeah, I, I mean, I think it's just a tool call. We can add that in. I don't think it's.
I would love that. Tell the team, please.
I know, I know exactly who to talk to to make that happen. Yeah.
But yeah, I haven't tried the use computer for that use case yet. My hope was that maybe dragging is supported now. But yeah, I agree, right? I think especially for us also like.
I don't think we tried it yet.
Like a task I've done, done a lot of work on is just export formats. But you need to test the exported file in some third-party software. And I have it installed on the VM for Devin, right? But then Devin can't open that. It can only use the command line version of these tools, which is often times limited, you know? So I'm really excited about use computer, even more excited about the Mac version that's in the pipeline. Because you know, again, our canvas service only be lined on web GL. And at least the previous implementation of the browser didn't have a GPU available, you know, which made the app super slow.
Okay. I mean, I'm sure these are things that the team want to hear from you. Yeah, you're talking to exactly the right person.
Yeah, no, I can keep on and on and on. We'll be another two days, you know, with all the, all the things I'd like.
But maybe here, back to what we're doing. So you see here, you know, it's chugging along. It's finding now all the components we need for the project. But now it's looking for the MCU that's Wi-Fi enabled and has a wakeboard co-processor. Right? Because the way this works, especially with a battery powered device, you wanna, wanna save power. And so you don't wanna leave the MCU running the whole time. Right? So simply how it's solved is that you have like a sub processor here that stays on, that's low power consumption, that just listens for the wake. But yeah, as soon as it finds the components, we're gonna see it, it's gonna start placing this. Right? So you can watch here as it does the job and we can talk to it and make corrections or changes. Right? And give it feedback. Any questions you have while we're waiting here?
I mean, it's interesting because like, obviously, this is much more of a difficult problem. But I wonder what you could do with pre-visualizations such that the center screen doesn't remain blank while you're doing this design phase. Like, how can I use this time while
the bot is thinking and creating to provide the next phase of input? I don't know how to do that.
We've been thinking about that too. I mean, a cool thing Rapplet does, at least for like a first user session when they sign up, is they use the canvas on the right side to show to like market features of Rapplet. Yeah, all the things you can do, right? And that was, and that's, it's like, I think for a first user session, that's great. That doesn't work so well, I think, for like, you know, somebody who's already on boarded. Another thinking we have here, right? As these agents here we deliver become more reliable and capable, maybe by default, we don't even show the canvas and everything on the right. We'll just show you this chat.
You know? And off you go. Another thought here is also like, look, right now this agent is as slow as it ever be. Right? I think you're now, you know, we're seeing the trend going also with the coding agents. You know, fine tuning open source models that are way faster as a result, way more accurate. You can actually beat frontier models on quality that way. I think a lot here, what's gonna happen in the next six to 12 months on our side is just like, you know, having specialized. Yeah, so it found an MCU here for us. We can look at this thing here. Yeah. Went for an ESP32. That thing is Wi-Fi enabled, has Bluetooth too, right? Has an onboard antenna, makes it really easy here. Don't need to have the designer onboard antenna.
Why are all the hashtags in the description?
Oh, this is for search. You can manually search here and so that way you can like say hashtag Wi-Fi, right? And you'll just find, you know, a lot of Wi-Fi related stuff. So it's just, you know, something users have adopted here.
Yeah, the renders are very, very pretty.
Thank you. Yeah, you know, a lot of this is also like you need to like see what it looks like.
Right? I like, well, how big is this thing actually, you know?
Yeah, you get a rough idea of this business. Yeah, and I noticed that you have a bit of a social network, right? Like you have usernames, you have followers and you have like the usage graphs and all that.
Yeah, everything is user generated, right? The whole library is user generated. Again, the spirit of GitHub here.
Hmm.
Why does everybody have to like maintain their own
compression library, encryption library, right?
Extract emails from strings library. If we could just have this joint repository that is GitHub for code, you know, to, to compete on the best implementation of a given thing, but then also distribute that and after feedback loop, right? And so it's the same thing here, right? This representation of the ESP32, right? There's ergonomics to this, how good is this 3D model? Is the wide data sheet linked, right? All this kind of stuff, that's have to write part number. And, and so you want this like friendly competition here about building the best representation, the most ergonomic representation, right? And so that's why the user-driven approach here has a huge advantage. Plus, you know, if there's something missing, users can just add it. They're not reliant on us.
What are we looking at here now?
Yeah, I was just gonna say, so it added, it's found an MCU with a wakeboard processor, added that. Now it's looking for the audio path components, right? So the speaker, the amplifier, right? You can see it's looking for class D audio amplifier. Oh, it seems like it's found one.
Yeah, this is a lot of autonomous thinking and searching.
Yeah. So yeah, here's our speaker. This is our amplifier here, right? That I just added. And what's this? Oh, this is the microphone. It's an onboard microphone.
And typically, do you buy these parts separately? And then a third party or middleman assembles them for you before shipping it?
Great question. So I mean, there is like, there's literally 10,000 of manufacturers out in the world.
This is like the Shenzhen markets.
Yeah, yeah, totally. So I think a popular option is in China is GLC PCB. They'll take the design out of flux and they'll print the actual PCB board, right? And then they assemble it for you and they send it back to you. Right? And that's a popular option because it's easy, it's cheap, it's fast. Like within a week, 10 days, you have it back, right? And you can have 20, 50, 100, doesn't matter. But you know, some people also, they want to solder it themselves for whatever reason. So they just get the board printed somewhere. The components themselves from a distributor and then solder it themselves. Yeah.
Well, I think you're probably coming close to this project being done. I just want to think about the AI engineering here, right? You said you use Langgraph, a Langchain. How do you iterate on your agent, right? Do you have evals? Do you have A/B tests? Or do you vibe check? What's the rough process?
All of that. So I mean, we got here essentially layer by layer. Right? So we built first the CAD tool essentially. Then we built like knowledge retrieval agents on top of that, right? And then from there, we went into two calls and to be like more sophisticated knowledge retrieval. And we started having evals for this kind of thing. So this is like foundational layer kind of stuff. And it's always a little bit of a mix of like evals and just vibe checking. Because in the end, it's not enough for the evals to pass, right? What's the legend here? That's the response, is the response to the both, right? Is do I have all my links in here, right? What's the information presentation really that matters, right?
And so that can stop with do a lot of vibe checking, that's trying it over and over.
But for things that are modelistic, like look, we need to find it easily to the library. Right? For that, we have evals, right? Uh, that we run. And so, but then on top of that, then we then started building these agentic features. First was, you know, it could perform single actions for you. It could add a component, remove a component, make a connection, right? And then again, for that, we have evals to harden that. And then once we had that, like, working, we started building the actual like agent now. This long running agent that we're looking at on top, right?
And for that agent, to be honest, that's mostly vibes right now. There's an increasing amount now of evals that we have, right? But I think how the V1 here was built, or the V0 of that was built was all vibes. Because you're just moving faster. I think the problem with evals often times is like,
Very common.
If you're early and you're still changing the architecture a lot, right? The evals can get in the way. It's a little bit like unit tests sometimes, right? Or like regression tests. If you're like still innovating a lot and changing a lot, you know? It's easy for like, uh, uh,
It's too early.
Like, yeah, yeah, it's too early, right? But yeah, it's kind of like, you know, it is very similar that sense to like regular code writing or like deterministic code, right? And that regression tests are nice, but there's all these typical issues with them. But they take a long time, they're flaky. And I think evals are very close to like where regression tests are. They take a long time to run, they're flaky, they need a lot of babysitting. One of the major challenges is surprisingly prompt management, call it. But there's so many like sub-agents here and sub-sub-agents, they all have their own little prompts.
And managing that and then also like even like, we're getting to a stage where we want to reuse blocks of prompts and other prompts. And you can put this all in code, of course, right? As like text files into a Git repo.
What we found is like that, you know, for our vertical, we have like domain experts on the team here that are electrical engineers, hardware engineers. And we needed, we need a way for them to easily improve these prompts, right? And the turnaround of like committing, like opening a PR, landing that PR, it being deployed is just too long. So most of our prompts live outside of the Git repo.
Where?
In Langsmith, actually.
And we're pestering, you know, them over there to implement like a concept of blocks so that we can have more reusability, right?
What is a block?
Isn't if you, if you think about like a prompt and it's like, I don't know, fifteen paragraphs.
I see. So sections of a block, uh, sections of a prompt that are reusable. Got it. Got it. Got it. It's, it's interesting how like all these are generally re-implemented everywhere. Like, yeah, there's a lot of like ginger templating. There's a lot of like just concatenate strings. There's different ways to concatenate strings, basically.
Yeah, yeah. We're trying to find like the one way to rule them all. We also have a bunch of hardcoded strings in the, in the code, of course, right? Well, like we, we don't change them a lot. It's fine. It's just, it's the hardcoded. Would it be nice to also be Langsmith? Sure. But nobody has the time to migrate it or whatever, right? But the ones that we work on a lot, they're all in Langsmith right now. And, you know, there's, there's some great things about that. I mean, we love the sandbox there. We use it for tracing too, right? And it's nice to be able to like open a trace, go from there to the sandbox, make some changes to the prompt, we run that, right? I think that's phenomenal, right? A formal UX point of view for us here. But yeah, the, I guess the CMS for the prompts itself, now we've been poking them there to invest more.
It's basically a CMS, right? Like it's, it's a very interesting form of CMS with like transclusion or like reuse and templating so that you don't, you don't mess up the format and all that. I, I think it's an interesting problem. I would say that the prompts management startup category has not been that promising. You know, I think there's like prompts layer that has been there. Human Loop got recently shut down. The company, we got like hired by Anthropic. Lang, uh, Smith has that. Brain Trust has that. You know, it's just not super clear to me where, Incognition, in fact, we have an internal one that was built. Basically, the same time as Devin, the original Devin.
There's an internal playground that was built that iterates on Devin's prompts. So the pattern that you found with Langsmith, they independently invented two years ago for building Devin.
And you have to, right? Because it's all about iteration speed and being quickly able to test new things, right? In Langsmith, we use the, you can label your prompts and we use that a lot to like deploy a prompt just for like a single user, right? Where like hacked our own system around that, right? But it's very obvious that a better CMS is crucial, I think, right? But also the reason why the startups haven't done well so far is I don't think we understand the problem well enough yet.
To abstract a generalizable solution, right? And I think that's where everybody is stuck, right? Even with Langsmith's office, there is like, you know, we use it, we've not found anything better, but it's so early. But I think it'll happen. Somebody will figure it out. And I think you need to like, need to connect this with the traces.
So that you can go like, you know, seamlessly back and forth, right?
Yeah, it's one seamless tool.
Yeah, and you need to connect it also to like split testing, right? So that you can deploy a change just internally, just to yourself, to 50% of the user base, right? These things all go together, I think, ultimately.
So you use Langsmith for all that? Do you have to bring anything else to handle this?
Yeah, for split testing, we use here, launch darkly. And there's a lot of overhead to configure a split test for a prompt, as you can imagine, as a result. So it's not great. And then, yeah, we use tags. Like I said, you can tag prompts in Langsmith. And then we have an internal system on a syntax for these tags to say, with this tag, this is a user tag, we're only going to deploy this prompt to this specific user, right? Internally or just to everybody internally for us. Or no, this is public now, right? This is the production tag. And so that's a kind of like, you know. Better than nothing. It's not great, you know?
Um, so yeah, I think it's just because to show that I think a lot more here to to be figured out before somebody I think can successfully build other the the prompt system as to rule them all.
Yeah. Um, okay, just checking back in on this agent. Um, how far? Is there like a progress bar on the show on the plan?
There's the progress bar. We've been thinking about that.
It's like an interesting problem of like, you don't know, it's like a, it's like a, you know, stopping theory, right? Like when we have no idea how to predict when the agent will be done, but we have some idea of like, how much of the plan have you completed, right?
Yeah, so I think the way we're thinking about it, we haven't shipped or built anything here yet, but we're thinking about it's like you have here the the plan. And we just want to show what tasks here have been competed. We can sort of like manually do this, right? It's still adding components. So we're very clearly here still in the um uh, ta ta ta ta ta, in the in the schematic development phase, right? Adding here the components, right? Right now it's doing the the power related stuff. So yeah, we're at step four here, right? Next up is B, then we will over to connect the components. So yeah, the the average session is, I think, twenty five minutes.
Wow, still, you're doing a lot of work.
Yeah, I mean, I think from a complexity point of view, it's no different than writing software.
Yes.
The the constraints are just different here. Like in software, essentially, you have like infinite characters available and you can always just write the function instead of finding a library that does it, right? And here you can't do that. You need to find these components and they need to be compatible with each other. And they need, each of these ICs needs supporting infrastructure. And you need to poke all this together. That's not to say that there isn't ways to make this faster and we're working on that, right? But yeah, the the first goal is always to just to make it work.
And then, yes, make it work, make it fast, right? Make it fast, yes. Yeah, so so I think like the
I mean, let's talk about like the sort of business side of things. What are the milestones that you've that you've reached that you can share? And and also, like, what is the the grand vision, right? Like, I think like one one thing that is unclear to me is, you know, is this for hobbyists? Is there a broader like industry use case that that you're going for? What is your sort of core audience that you're achieving?
Yeah, great question. So we have seven thousand paying customers today. That's like, you know, it's grown, it's like twenty six X versus last year. Right, so massive growth here.
And all over the place, like, yeah.
Yeah, so it's all organic, right? People find us on social media.
You have a very good video.
On your YouTube.
The team here does great work on on on social media content and videos and tutorials and all that. Google search is also a big source of users, right?
Is there like a big Reddit?
There's not a big Reddit, no. We, I think none of us here is a huge Reddit users. So as a result, you know, we're not big on Reddit. But but we have a community Slack that's pretty active, you know, we're we're hanging out there too with users. That's probably the main channel to to interact with users today. And then, um, what was I gonna say? Yeah, so that's, you know, most of users are soft serve. We have some enterprise customers already, you know, Fortune ten, five companies, right? It's starting to adopt it. We have a lot of like, we have a lot of individual engineers at large companies. Like, you know, if you look on your desk, the brands you see there, we have like, it's about these fifty, sixty, you know, individual engineers using the product and paying customers.
It's very like similar to like Notion or Figma's go to market motion here, right? Uh, the product leads here, hundred percent focus on improving the product and getting, uh, enabling users to get more value out of flux, right? And that's how I think about the go to market motion here. In terms of what's the ambition here, I mean, look, right, right now, we're focused on PCB boards clearly, right? But I think ultimately the vision is like, if you look at how you can prompt a poem into existence with ChatGPT, right? Or a piece of code with cursor, we want to be able to prompt hardware into existence. And not just to PCB boards ultimately, but the full thing, right? Like, why would you not be able to prompt a smartphone into existence?
Right? And if you can do that.
That's a lot of parts.
That's a lot of parts. Yeah, yeah, yeah. That's, that's, there will be a thousand, two hundred parts or so, right? For a smartphone.
For reference, this is how many, what we're doing here is ten?
Oh, this is probably gonna be like, when we're done here, probably like fifty or sixty or so.
Wow, okay.
And so look, and if you can prompt something like a smartphone into existence, then why would you even still go to Amazon and spend two hours searching for a product? If you can just describe what you want and get it made and sent to you within a week, right? And I think the counterintuitive thing is that we have been all led to believe that to make things cheap, they have to be mass manufactured. And that used to be true in the past, but that's been changing right through automation. And many of our customers now, right? The total cost of ownership now of the electronics, whatever they're doing, is like one hundredth of what it was before. Right? Because if you can have an AI agent produce the design for you, then you're only paying material costs to manufacture it.
And that's a fraction of what you pay for an off-the-shelf smartphone and off-the-shelf Alexa device, you name it, right? Of course, there's some categories that are, uh, uh, they're gonna take a long time to get to. I think like doing something like an iPhone is, uh, is gonna take a long time. Simply because the iPhone enclosure is essentially a piece of jewelry. Also, that's gonna take a while, but if you think about it, the majority of electronics, like when it's manufactured every day, aren't iPhones? Not even close, right? Like electronics are in everything you use, your headphones, your car keys, your charger, your microwave, your washing machine, your thermostat, your car, right? I mean, it's just everywhere. And so there's a lot of low-hanging fruit here for us to come in
and, and help these companies that make these, uh, uh, products, you know, customize them more, make them cheaper, make them more capable. And so we have a lot of like, we have like, you know, vending machine manufacturers use flux. Because electronics and vending machines, right? We have like a traffic light companies, because, because what? Traffic lights, right? There's a computer in there, right? We have farmers making their own farming and farm automation equipment with flux, right? So, uh, yeah, wherever you look, where there's electronics and automation, right? We're coming for that, you know? And I think this is the end for most categories of the OEM model ultimately.
This, this thing looks like it, it might be, uh, taking a while and, uh, uh, and like I do have to jump to the next thing. Um, oh. Is it done? It's done?
Uh, yeah, the work's done.
All right, then maybe we'll review it a little bit and then we'll uh.
Yeah, let's do it.
So how do you look at it? What do we do? What does the person do next? Yeah.
Yeah, exactly. Good question. So it's looking what it came back. It says, you know, it's done here. And all the key components. It wired all the key components here. Oh, yeah, we recorded some some estimates here on on battery lifetime. What is not done? The low power wakeboard component has not been added to this command. Okay, we need to go find that. There's some things it can do that we have to do manually here, told us. And success could, so then we recommend it next steps, right? Oh, yeah, we need to like do some research on the wakeboard co-processor. And then, you know, go to the layout here, figure out the antenna stuff. And then we can use auto layout, you know, for placement and to run the E R C D R C checks.
So that'll be the next steps here, right? But we got pretty far here, right? We got all the components here. We can now go in here like review this manually. We can make manual changes. So again, this is very similar to like, you know, what the workflow looks like with Devon or Cursor, right? The agent will get you to a given state until either it's done or needs feedback.
And then you can take it from there.
But no, I think overall super interesting and I think um. You know, like you are relaysing your vision of AI hardware engineer. I guess like the last thing is any call section on both like the, you know, China's outside and the hiring side.
Yeah, absolutely. I mean, definitely come, but if you're a hardware engineer or software engineer interested in hardware engineering. I think we're like dramatically lower to barrier to entry and speed to process up. So definitely come, check us out, provide feedback. Oh, this is like day one. There's a lot more work to be done and we would love you, you know, be part of this and help us out and direct us. And the hiring side, yeah, look, we're hiring. We're hiring software engineers. We're hiring hardware engineers. We also have a bunch of go-to-market roles open we're hiring for. And I think we're looking for people who are like, you know, excited about making hardware and democratizing hardware, excited about what about AI here? And really in inventing what the next generation of hardware design tooling looks like, and next generation of CAD tooling, what it looks like here with us. It's exciting times, you know.
Well, a lot of, there's even interesting technical questions around the canvas that you're doing. It's, it's so, it's such an interesting, challenging problem.
Yeah, yeah, it's a whole can of worms on its own. The whole GL rendering, definite performance in any browser, on mobile browsers, memory usage. I mean, you know, there's no shortage of fun here as an engineer.
Awesome. Well, thank you so much for your time, Matthias. I'm so glad to have like this personal demo from you on Flux. It's, this is a real problem that I've actually tried to solve on my own. And to see it solved live in like 20 something minutes is actually really compelling. Like, I think like you have a lot of confidence in your tool, but also like you have a potentially unbounded problem of building hardware, right? Building the iPhone, so.
Obviously, there's an infinitely large market here, right? And it's just like a capability that people haven't had at scale up until this point. And I think that's what's exciting here.
Awesome. All right. Thanks, Matthias.
Thank you. Have a good one.
Page updated October 24, 2026 · transcript and summary are machine-generated
