FoDES - Future of Design & Engineering Software
We discuss tools and technology that engineers will find interesting and useful. This can be software, hardware or a service.
FoDES - Future of Design & Engineering Software
Gustavo Navarro, Founder of Divergence AI - AI Copilots For RF Engineers
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
We explore how an AI copilot layers on top of HFSS to automate RF simulations without losing rigor or control. Gustavo Navarro shares a live demo, a practical roadmap from post‑processing to pre‑processing, and a vision for cross‑domain orchestration across trusted solvers.
• Why HFSS expertise is hard but essential
• Natural language to HFSS automation without hiding code
• Generating S‑parameters, 3D patterns, and full reports
• Interactive agents that ask for missing setup details
• Orchestrating sweeps and long runs with monitoring
• Using ML for fast screening, solvers for validation
• Pre‑processing: geometry creation and defeaturing
• Moving RF designs across tools for platform studies
• Roadmap: a future with thermal and structural workflows?
Meet Gustavo And The Mission
SPEAKER_00Hello and welcome to Fodez, the Future of Design and Engineering Software Podcast. My name is Rapindertara. On this show, we will have guests that will discuss tools and technology that engineers will find interesting and useful. Today's guest is Gustavo Navarro, founder of Divergence AI. Gustavo has left his native Chile to be come to the Bay Area to give engineers a helping hand with AI. By this time, you might be thinking, isn't everybody? But Gustavo is not some opportunist jumping on the AI wave. He has a PhD from E. Davis in mathematics, and his knowledge of RF design is a mile deep. So why should you pay attention? Because what Gustavo is doing with AI and RF, adding a copilot to ANSISHFSS, could serve as a model for all sorts of simulations. Hi, Gustavo. Nice to meet you.
SPEAKER_01Nice to meet you.
SPEAKER_00Welcome to the show. Thank you. I'm excited to be. I think we're, I wouldn't say neighbors, but we're reasonably close by. You live in San Leandro? Yeah, that's right. I'm in Marin County.
SPEAKER_01Okay. I go surfing in Marin County very often. Oh, you're a surfer? Well, a little bit.
SPEAKER_00I was looking up a little bit of your background, and you're from Chile.
SPEAKER_01I grew up in Santiago, which is like a bit old city in La Cena, which is maybe like a coastal town, if you imagine. Chile is like California, but upside down, right?
SPEAKER_00Upside down California. Really, really long, right? Really, yeah. A lot like California. I think you share a lot of the probably the surf, but also unfortunately, the earthquakes, right? Oh, yeah. I think the worst earthquake ever recorded was in Chile. One of the worst.
SPEAKER_01My wife says that I'm an earthquake snob. I don't flinch for anything above five, you know. Anything below five, I was like, oh, that's just a little tremor.
SPEAKER_00We had to get used to that ourselves. My wife's from Philadelphia, so about earthquakes, but I'm from India, right? And India up in the Himalayas is a huge seismic zone, like one of the biggest in the world. I'm kind of an earthquake snob myself. Exactly. I see you have a PhD from Davis. Do you find that people stop talking to you after they hear that?
From Chile To Math And RF
SPEAKER_01The classic response is oof, I hated mathematics when I was growing up. And I was like, Well, tell me more things that you love so I can say that I hate them too. I always try to use the opportunity to convince them that mathematics is kind of beautiful. I think they have been kind of misled by the typical mathematics propaganda of school, you know.
SPEAKER_00I try to tell you you don't seem like the kind of person that would scare me as a math instructor. So what examples do you use to make them?
SPEAKER_01My classic example is that what you learn in school is people tell you that you're learning how to paint, for example. And you're like, Oh, I'm painting it. And yes, you're painting a fence, right? And you're painting a fence and it's white and everything else. And you practice a stroke and you learn about moving the brush up and down. But would you call that art? No, right? Like you don't really feel, it doesn't feel like art. Therefore, you shouldn't call it mathematics either, right? So what we're talking about in math, you're thinking about Picasso or Michelangelo. That's really painting, that's really art. And so that analogy reflects into what they're thinking. Like, oh, maybe I didn't learn math. What I learned is some techniques to do some calculations, but the beautiful part of it I didn't really get to explore. That's the analogy that I kind of tell them. There is a beautiful world out there of geometry and proofs and the exploration of the physical world through mathematics that I think they're missing out. So that's my typical approach, and hopefully I get them.
SPEAKER_00Does that work? I think my wife was impressed enough that I was an engineer without having to learn what engineering really was. How'd you ensnare your wife? Or is she a mathematician?
SPEAKER_01No, she's not a mathematician. She was like, Wow, that's such a romantic view. Of course, we've been together for more than 10 years, and so she knows that that it is hard and you have to think a lot. She's not so keen to study, but at the beginning, she was super, she bought it up. Yeah, I imagine, intimidated, right? Because that's not no, she she I would say that she's very smart on her own way. She's a PA, like a physician assistant, she's very scientific herself. She's also very funny, so she gives me checks.
SPEAKER_00So she's never like, oh, my husband. I promise we'll get back to engineering soon. So my wife is an English major. I'm an engineering major. So that was a complimentary thing. Now, your your wife is a physician's assistant, so she studied medicine. Do you have this argument all the time? What's harder to learn? Medicine or engineering or mathematics in your case?
SPEAKER_01Early on, I think we had this argument more, but because for example, Jesse had a profound admiration for medical doctors. Okay, because their job is very serious. My job I chose almost specifically for its distance from the real world. Although it turned out in the end that it was not the case. When I was young, I was thinking like there's this mathematician called Hardy, the spiritual mathematician. He would abhor if his mathematics were used for something useful. He would do everything for the sake of just intellectual pursuit. I was the same. I was like, I don't care if this is useful. So this there's this big stark difference between why she does what she does, which is to be as helpful as possible, as practical as helpful as possible, and why I do what I do, which is for the intellectual pursuit, the beauty of it all. Like even if it's not useful, I would still love to do it. And I think that was our main challenge, not in terms of difficulty. Of course, medicine is hard, takes a lot of years, and math is also well known to be hard. So there was not that different, but how useful it's going to end up being yes, right.
SPEAKER_00So my daughter went into nursing. I couldn't talk her into going into engineering. I was very unsuccessful with convincing either of my children to go into engineering. So she studied nursing. How deep and she's always trying to tell me how important her job is. I said, We can I can kill thousands of people at a time. You can only kill one. Yeah, I mean, exactly. Our impact is measured in the hundreds of years, right? Yes, right.
SPEAKER_01I have a young son, so any tips to convince him to become a mathematician or an engineer.
The Beauty Of Math Beyond School Drills
SPEAKER_00So I'll give you some advice. Took my son to my office with me, trying to impress him with engineering. Look what I can do. Look, I'm he specialized in computer-aided design. Watch this, son. And he's sitting in it, he's watching for a little while. Eventually he goes and sits in the corner, totally unimpressed. He says to me afterwards, Dad, this is what I want to do for the rest of my life, is just and he makes a motion of pecking at a computer all day. And then you know what happened? He ends up using the computer all the time. Or composing, yeah. I do want to hear about what you're doing. It's very impressive. I found a little bit about it. You may have to talk down to me. I'm just a mechanical engineer, right? Oh, happy to explain anything, actually. But give you my background so you know what level to talk to. Tell me what company does. So Divergence AI is really narrowly focused on, correct me if I'm wrong, on H HFSS. I barely know how to spell it, right? HFSS, which is an application assist for antenna beams, correct? Antenna beam, antenna wave propagation, right? So what does your software do that helps with that?
SPEAKER_01We are building an identical pilot for electromagnetic simulations. If you're a hardware maker that builds a communication device, a jamming device, a radar, or anyway, a ton of stuff has communicates, has antennas actually. Or or RF is called radio frequency systems, then you probably use one of five most popular software out there in the world, ANSISHSS, which is high-frequency simulation software, being one of them, one of maybe the leading software part of it. And so our copilot helps you automate your workflows with HFSS. So it's kind of giving the power of coding to engineers that typically don't have the time to actually dedicate to create automation. So they do most things manually.
SPEAKER_00Okay. So now I've used ANSIS before, but the structural stuff. You have to spend a lot of time, you have to learn its terminology, you have to almost become an ANSIS engineer to become an FEA engineer. And I imagine it's the same way with HFSS, correct?
SPEAKER_01Yeah, that's absolutely right. The learning curve is very high. So what ends up happening is that companies have RF teams, they are very sophisticated engineers. They have a master's, they have PhDs, and they are the ones that run the simulations, build the designs, and train other people. So, like you say, it's a very sophisticated tool. It requires a lot of expertise. It's very hard to use.
SPEAKER_00If I'm using divergence AI, do I still need to know how to use HFSS? Is it something that helps me automate the usage of HFSS, or is it something that helps me not learn HFSS? I can get my antenna beam patterns without knowing it. That's what I'm wishing for, but that's a great question.
Why HFSS Mastery Is So Hard
SPEAKER_01So, okay, so I'll tell you where we are right now and what our vision is. I'll say right now, you still have to know what you're doing. And I think engineers like to know what they're doing. I'm an engineer too, like I'm a mathematician, but I've been a physicist and a mathematical engineer, as they call it in Chile. And I've been working with engineers for a long time. I think we don't want to relinquish control, right? So we want to know what we're doing, especially for the hard bits of simulation. The whole goal of what you're trying to do when you design something in HSS or any simulation tool is to know how it will behave in the real world. You want to know that answer as close to reality as possible. You need to make sure then that your setup, like your materials, your geometry, like your fabrication tolerances, like you explore all these things to the 98% uh accuracy at right. And then you run your simulation and you control the mesh size and you do all these things so that when the results come back, you can tell your boss or yourself that listen, now we're gonna fabricate it and it's gonna perform like this, and I know this for sure to 98% accuracy. If you don't have that confidence, then what is the point? You're still gonna spend time building it, it will fail the test, and you have to go back and redesign it and test it again. The whole point of simulation is to skip that. On the setup, you still want to be specific and careful. Right now, our goal is to lower the difficulty levels. So maybe more junior engineers that are RF trained, maybe masters or they're just graduated, they can still use HSS. More than that, they are able to automate to create more complex workflows than what they can currently do using our tool, but it will not replace their thinking. And I think that's a fear that a lot of engineers express to me, which is that they don't want to be dumbed the tool down. They don't want the tool to be dumbed down. They want it to be, I don't know, maybe, like I said, easier to create automations, perhaps, but they don't want it to replace the thinking. And I obviously don't want that. So you're reducing the complexity and its usage, but you still have to know what you're doing. The vision for it is to be an intelligent co-pilot, right? So it can give you suggestions and it can tell you, would you like to set it up like this, for example? The rest of your team at your company set it up like this, and the reason is X. Would you like to do that? And you can say sure. And they automatically fill in the details if you want to, but you will always have that being to say, no, no, no, I don't want it like that for this case. I want it like this.
SPEAKER_00Got it. Correct me if I'm wrong. This would be like the senior analyst here who's doing RF RF a lot and knows HFSS. They would help with the automations or tell their junior RF engineers, here's how you can use divergence AI for what you're doing. By the way, you can do it a lot faster, and you can do many more cases. I really love the demo. I really love the natural language interface you have with that because you can tell quite clearly something that would be difficult to learn. Steer the beam from zero to 30 degrees and then give me the beam patterns every 10 degrees. Tell that in English.
SPEAKER_01Yeah, and that's a great example that you're bringing up, which I have a demo prepared if you want to see it. But the I would like to.
SPEAKER_00Are you able to show a demo?
SPEAKER_01Yeah, yeah. Should I do the array? Maybe I'll do two projects. One is more like a 5G waveguide project to see some post-processing capabilities and for creating reports. And then we can do one with an antenna array, which is more for people building radars or satellite communications or 5G wireless communications stuff. Okay. And then if you want, I can show you a teaser of our latest stuff that is under development, but only if you want to. Okay, yeah, please. Okay, let me see. Can you see my screen okay?
SPEAKER_00I can see your screen.
SPEAKER_01Okay, perfect. On the left hand side, you see the ANSISHSS software, and on the right hand side you see our software.
SPEAKER_00So you get all your geometry from it, it's already created.
SPEAKER_01Yeah, so we're connecting to an existing project. You basically want to do automations on top of this project. This could be something that a designer is building, for example, this little array of this little kind of square elements that resonate with the waveguide underneath. This radiates like a beam type thing. The engineers might build it, and then they want to do things with the design, whether this is modified for a specific requirement, maybe a resonance frequency, so it's more efficient, it's more efficient at one specific frequency, for example, or they just need to create a report, or like you said, some of the things like plot a beam and move the beam around. On the right-hand side is our software. Our software has three components. On top is how you load up a file. You browse the file and click load. Our agents can write a little bit of code to connect to the design and extract design context. This is the key part of our work, actually, is to extract the design variables, the ports, the basically everything that is related to the design itself, so that the agents now have this awareness. And they write code on the right-hand side, which is our Jupyter notebook live views. Every piece of code that the agents write gets written into this notebook that the user can then use later on as well. We believe in giving all the control, like facilitating the work, but then giving it back to the user. So we don't try to hide the code that was generated.
SPEAKER_00So natural language sort of interface on the left.
What The RF Copilot Actually Does
SPEAKER_01Yeah. Code that answers HFSS can understand. On the right, exactly. And results are also on the right. For example, you can see that when I loaded this file, it's successfully loaded. This file has three designs. The summary is that SIW is solved, the 0.5G element is not solved, and the two SIW structure is solved. It has two ports and 25 variables. Now we can start analyzing and asking questions directly for us. And you can ask things like plot the return loss for all active designs. This is a classic thing that any general wants to know, which is the response of this antenna at different frequencies for all solved designs. As you mentioned, our agents basically take it like a language request and they're trained to know what code to write to interact with the design on the right-hand side. You can see that you can look at the execution trace and it's using a few tools that we have built for. And something is going on on the UI. Our agents are talking to HSS and writing some code, and the output is coming in here. You can see this graph is the frequency response of the first SAW structure of the waveguide. It asks you, do you need a comparative plot for the other design? So I'm gonna say yes. That's nice, it's interacting with you.
SPEAKER_00Yeah, yeah, exactly. That's what I've got so far. And if you want anything else, you want some changes, or does it sometimes ask for you to refine your question if you don't have something that it understands? Yeah, exactly. So sometimes it would write code that fails.
SPEAKER_01For example, it says if you ask it, I don't know, plot the radiation pattern, and it turns out there is no setup associated to that. And it's gonna try to do and it's gonna fail. And it says, hmm, it failed. Maybe it failed because this is not there. We let me give you instructions of how to set it up. It will help you through that process.
SPEAKER_00You forgot to model it, you dummy, something like that.
SPEAKER_01Yeah, exactly. Sometimes it says there's no you are asking for things that have no results yet, and that definitely happens, yeah. Yeah, so I'm gonna say plot the one for the two SIW design because I want to see the response of the full with the one with the array. The agents got to take a look in here and uh did it plot it already? Oh, yes, I plotted up here. Sorry, it's a dummy, I already plotted it. It's up there, take a look.
SPEAKER_00Taught it to be polite, good, all right.
SPEAKER_01Exactly. And so you can you can keep asking questions that for example. I think that engineers kind of care about is like about the speed chart. Our agents also know how to do that, and it allows you to understand what is the matching, like how good is your match to the frequency that you're trying to do. And sometimes when the engines need to think, they create this to-do list. So it has this ability to say, okay, if I'm gonna set up a long workflow, maybe I'm gonna break it down into pieces and I'm gonna attack each piece one by one. Here you go. This is the speed chart with the impedance for s11 over these frequency ranges. Obviously, this is very manual still. If you were to do it, you still click, and right now this feels not as a big advantage, but the advantage goes into saying, okay, maybe I want to create a full report. So you can say, okay, plot. I'm gonna say plot s11 for all designs, then plots. Maybe okay, there's another Y11 for all, and then I say then plot the far fields. I don't know what else we can do. We can do more stuff and then plot the B.
SPEAKER_02Okay.
SPEAKER_01Plot the B. Maybe this is you describing the beginning of your workflow, like everything you need to be happy, and just letting it go and letting the agents do their job. You can go do some tests on the bench, like you create this to do list, and as it executes, you can be free.
SPEAKER_00You can even ask to this to do list, say notify me by when this is you could actually run hundreds of simulations, sequential simulations, and just walk away and it works for you.
SPEAKER_01Yeah, and I think this is where the power of what we're trying to build happens because normally what an engineer will have to do to build this automation is that they will have to be very good at writing this code right here, this type of code, which is using the ANSIS Python PPI.
SPEAKER_00That's a very rare skill.
SPEAKER_01So those people are like unicorns.
SPEAKER_00It just scares me looking at that stuff. I can't do that. I don't want to devote six months of my life to learning how to use that code either, right? So this is worth its weight in gold, something like this. Tell me how much gold is out there because I don't know much about this specific field, but it's gotta be more than dozens of people who are doing it.
SPEAKER_01So we are going after the big industries out there. We're going through aerospace, telecommunications, consumer electronics, wireless design.
SPEAKER_00Yeah, wouldn't 5G, all the antennas they're making now, really benefit from this sort of simulation and automation? Exactly.
SPEAKER_01So our ideal customers on that space are Cisco, for example. On the consumer electronics side, you have an Apple. How many antennas are in the iPhones? I think it's something like 15 antennas.
SPEAKER_00Exactly. I was just gonna ask you about that because one of the problems I read about in its development was how the antenna is going to fare if it's cross the top or the side. They had to spend a lot of time with that. And I don't think they had your tools, Gustavo, right? They were probably doing it iteratively one step at a time, trying to figure that out. That was a tough problem for them to solve where to put the antenna and not have it not drop the call, but depending on how the phone is oriented.
SPEAKER_01Yeah, it's all radio frequency, it's all basically antenna antenna engineering.
SPEAKER_00It's very it's all this applies in several fields, not just RF, but making the invisible visible. Yes, one of the main things about simulation software. Another application I'm thinking of, tell me if your software can help with this or your AI can help with this, is a vehicle-to-vehicle communications, which uses uh I think planned on using 5G, but that must have a lot of antennas, right? Auto autonomous driving. They have a raise of antennas, they're all good applications.
Live Demo: Natural Language To Code
SPEAKER_01Yeah, exactly. And so actually, this is a great point. You're making all the advertising for me. My goal is to help engineers when I talk to RF engineers. I consider them to be the most brilliant engineers, by the way. Sorry to the mechanicals out there. They're very smart. But they always say RF is black magic or incredibly hard, and I agree.
SPEAKER_00But it's it's also very beautiful. So you have a mind that can see things I cannot, right? To me, it's like I can't see it, it's not intuitive. You seem to be able to see it, but better than that, you can explain it and even help analyze it. When I talk to them, I try to think of what problems they have.
SPEAKER_01Obviously, they always complain about simulation time. Simulation time is too much, and there's a lot of companies that are trying to solve that problem with AI, they're building physical models. I think you end up being podcast. And so they're, you know, their building is super interesting models. Right. But there is other times, right? That there is a setup that is very manual, there is repetitive tasks, there's like workflows that are not well suited for automation unless you dedicate a ton of time. These engineers are like very brilliant, they don't have time to do this, to this dedication. I'm sure they could learn how to code and do it. This is certain, but they don't have the time or they shouldn't dedicate it to this. They should be thinking higher-level product decisions. One of these examples is you design an antenna in HFSS, for example, and then you want to put it in a car or an airplane. The process of moving it from there to the car involves separate software, actually. Let me do a plug for some of my friends here at EMA3D, for example. They build simulation software to understand the propagation of the antenna fields on a car and see how it affects the cables or into an airplane. That workflow is very manual. The engineer has to take the design, export it to some file, put it to another thing, run the simulation, get a report, and then finally understand. Whereas we aim to be that orchestration layer that you tell the agents, do this. This is not a heavy thinking process. And the agents say, Okay, writes the code, takes the fields, move it to the next file. It now interacts with the other software, gets a report, and send it back to the user. And the user can say, Oh, now I see the outcome. Yeah, our goal. We're working hard to integrate with those parts.
SPEAKER_00It's great that you're building it on a foundation of tried and trusted simulation software. You're just adding a layer of AI on top of it and by oversimplifying. This is much better than other AI tools, which we're trying to synthesize the simulation. Go to the next step in an analysis by similarity to another analysis without using the code. There was something very scary about that. I understand you can synthesize if it's very similar. For example, you change the wing design a little bit, right?
SPEAKER_01Yeah.
SPEAKER_00Maybe you can extrapolate to that, right? But if you're gonna change the wing design a lot, if you're gonna go from box body, your extrapolation is not gonna work. You still need the code. You still need the simulation code.
SPEAKER_01Maybe before I answer that question, let me just review the outcome, Aitan. The entire list of the two plots basically are here, and uh the Y1 one also plotted it here. It wrote all the code, and this is the 3D pattern of the antenna that we just plotted, and then finally 2D cuto it. Okay, good, good.
SPEAKER_00So you're all I've seen before is beam patterns, but you can do the full 3D. Yeah, full 3D also. Okay.
SPEAKER_01And the point was that we just left the hub and we didn't have to do anything. You and I just chatted, right? And you suddenly have this report ready to go, which is the goal to release the engineers to do more stuff. Let's go philosophical because I like this question a lot. I posted on LinkedIn about it where I was thinking, in my head, I'm a mathematician. I studied the algorithms that the solvers use behind the scenes to model the physics. So this are is a differential equation, it's an approximation of reality, right? Like the wave equation is an approximation to some degree that is accurate. Right. We are discretizing it into a mesh, and then we are propagating these discrete values of this differential equation through this mesh. We iterate it over time. And for all these steps, there are errors and we accumulate errors along the way. And we are just happy to know that the mathematicians of old and the physicists of old found bounds to maintain this error small enough. Like we discovered that if the mesh is small enough, then our error is small enough. There's a proportionality there. So we're kind of happy that we trust it so much, it has a history of being effective, and so we trust it. We say, Oh, finite element has been good for us building bridges and stuff for the last 50 years.
SPEAKER_00Yeah, we're maybe we may not like working with it, but we like it. We know it works, it's a solid foundation, right? Don't throw it away.
Reports, Sweeps, And Beam Patterns
SPEAKER_01So I had a philosophical question about the AI models. If you just look at them abstractly and say, this is just a method, right? Forget that it's AI or machine learning. It's just a method that allows you to approximate the function, which is the solution, right? What's the difference? If we eventually have the scientists telling us the error bounds, and we know that the errors are small in comparison, distance to the original training data. Like you said, if you start with a wing and you take a box wing, maybe the distance from the new data to the original training data is too much. And so the error is gonna be big in comparison. But if you're doing modifications that are small in size in comparison with the training data, philosophically, what's the difference? I tend to see these two methods as just methods of solving it. Yeah, I don't consider them necessarily bad.
SPEAKER_00Two methods, and you're right. If it's either small time extrapolation or small time interpretation, it can work. That works really well with uh rendering, right? Visualization rendering, where you can see, oh, this is gonna be that color in between these two colors. So fill in, fill it in. Don't wait hours for the computer to do the ray tracing, fill in that last little bit. Also, it might work if the model is diverging to a solution. Maybe AI could step in and say, oh, the final solution is gonna be along this because this is asymptotically approaching that. Yeah. But please don't make the big assumptions, don't throw out the engine and expect the car to run. Yeah.
SPEAKER_01My opinion is that the AI algorithms for physics will have a very important place to play in this whole orchestration. And they're biased to think that my orchestration tool will actually help engineers move designs in the right cycles of AI design and then move them into validation stages with the real algorithm. These systems will have some tolerances. Maybe they will give you 70% approximation and 80% approximation of the of how we'll behave in the real world. And perhaps that is good enough early on. You might want to trade off accuracy for speed, and you might say, you know what, for this antenna, run it 10 times or for 10 different designs on my AI engine, give me 80% there, and now send me to a validation stage. Of course, our agents are going to be the one moving this thing around, you know.
SPEAKER_00I did notice that your agent is a little bit more interactive than what I'm seeing. I see a few of these products, and most of the time they just spit out an answer. Sometimes it's good, sometimes it's way off. But at no time do they say something like, wait, I didn't catch that, or you didn't model that. You and your software said, Okay, you didn't give me that information. Now when you give it to me, I wish AI would be more interactive like that. All AI should just don't just come with me and answer because you give me the wrong answer, I will I will probably not use you again. If you work with me, right? Yes, if you sense where I'm going and maybe I need to give you more information. I love that, right? As an engineer, design engineer, I love that assistance. They help me with the tasks that I have trouble with, but you know, learn from me. Don't just jump to the answer. One time it was a kind of a joke where there was a sign draped on a building under construction, and they said, Hey, AI, finish with the building. The joke was that AI can't really make buildings, but there's enough AI right now, especially this prompt to shape software that's trying to do that. And as an engineer, I'm extremely worried. Skeptical, yeah. Yeah, that people will say, Oh, that's the AI I want. I don't really need the math. I don't need the structure, I don't need the FEA.
SPEAKER_01I think that's very interesting. But what do I feel about it? That's a great question. Obviously, I'm looking at the space, it's important for us. A big problem for RF engineers sometimes is something called defeaturing, which is we have that in the mechanical side too.
SPEAKER_00Exactly.
Where AI Helps And Where It Shouldn’t
SPEAKER_01Because for maybe they don't matter for thermal analysis. And if you were to mesh every single detail, it would be just a smash, which means your normal simulation time, and maybe it doesn't matter. And as an engineer, you want to know that you're playing that trade-off, right? But you need to know. I think that's the key. Yeah, and so for us in our copilot, what we're seeing in our roadmap is that we want to help engineers do the pre-processing. For example, if you're putting an antenna on a rocket, one of our potential customers wants to sometimes they get the rocket from the mechanical team, and it has so many little parts that they are like, forget the rocket. We're gonna draw our own rocket, they draw a cylinder, they slap the thing. The reason for that is to play this trade-off, but that process is so time consuming that they prefer if some AI can take the original shape and simplify it. We are working on that right now. We don't have it, but we're trying to work on what we're calling the pre-processing agent. What I showed you was a post-processing agent. The simulation is already run and allows you to we have a simulation agent that allows you to run optimizations, parametric sweeps, and it monitors the status of your run. If it fails or if it's not converging, the agent will jump in and say stop or run again, just in case you let leave something running overnight, and then you don't want to wake up the next morning to a fail simple. So basically, a simulation agent is like having a junior engineer. Now we're working on the pre-processing side, and and this is maybe some of the things that I was thinking of showing you if you're curious of what we're trying to do. Yeah, please. If you have time, I'd love to do actually. This is under development, so maybe it will not work right away. This is running on our development branch. So this is an empty project, and what we're trying to work on is giving the agent the ability to draw some stuff. This is important because for RF engineering, a lot of things are out of the box, like a patch antenna, for example. This should be easy, a dipole or a horn antenna. These are things that you use for testing that you don't want to be porting it all the time or drawing it from scratch all the time. So we want to facilitate that process. So we here I'm giving the agent the ability to create some geometry. So let's just ask what types of antennas can you create? And of course, our agent has a list of things they can do. Let's see what we get.
SPEAKER_00So it's thinking, it's thinking.
SPEAKER_01It says I can create a variety of antennas, okay? A bow tie, standard, rounded and slot, patch antennas, pro fed, and horn antennas, and helix antennas. Let me know if you like details of which one and what are they for. So maybe we can ask. So, okay, so let's just create uh micro strip, micro strip patch antenna with a instead feed. And we have to tell it the frequency because that matters. Let's say Wi-Fi frequency. So we used to build this patch antennas all the time in my old company where at Breach Power. Uh we can still build an antenna arrays. You can see that the agent is writing some code and actually is creating the geometry. It added the geometry right away and created the antenna. I don't know. And the whole goal is like this used to take, I don't know, an hour. You can just ask it and get it done. And we're working on creating arrays now. So you can say give me an array of five by five of this boom boom boom and have it to be reusable.
SPEAKER_00Okay, now this was all done by reading all the documentation for the ANDS software.
SPEAKER_01And then yeah, actually, so this has a very good open source library called Antenna Toolkits. What we did, we looked at this and said, okay, this is perfect. I started asking our F engineers to use Antenna Toolkit, and they're like, no, they build these things from scratch. And I'm like, why do you build it from scratch? And it's because it's hard to use. It's literally the only answer. So we're like, let's just make our agents use it instead. And we integrate it. And this is the same thing that we're doing with very brilliant engineers from Manitoba University. They built a library to analyze wireless networks, but it's a Python library. So that means that if you want to use it, you need to know how to code. You have again this dilemma of RF engineers that if they want to use very advanced Python, they have to code it in. We are collaborating so that we have an agent which is a wireless agent now.
SPEAKER_00I love it. I love it. I want to do the engineering, you do the coding.
SPEAKER_01You do the engineering study. We do the coding and we give it something back. So for example, now it's saying, okay, I make this rectangular micro strip antenna and I use this type of substrate for which this thickness, this the dimensions were automatically calculated for that resonance frequency, and the setup is ready. So we like to run a simulation. Let's say, yes, please, let's run it. So our agent should and you can see that it's happening here. So our HFSS is starting to think and it's loading, it's running this analysis. We didn't really have to do anything. So once this is integrated to our main code, you will just be able to set it around and get out of here, and you'll have the answer in I don't know, five minutes. Or so and little by little we want to add more power to the missing agent. This is gonna take a while, so maybe we can leave it thinking while we touch.
SPEAKER_00I'm gonna think selfishly again, but mechanical engineers, right? Couldn't there be something like this for stress analysis or vibration or thermal? I used to do that stuff, but years ago, but I doubt if I could do it now without having to relearn it, right? So here I am asking for is there a Gustavo and the structural side, right? Is there are you gonna is that the plan? You gotta try to do this or discipline. You're asking the multimillionator question, right? Sure, yeah.
Orchestrating Multi‑Tool RF Workflows
SPEAKER_01Yeah, yeah. We we want to solve. Uh so I fairly new to the valley, right? So I am I don't have a big track record of companies. One of the things that I want to do is make sure that we have a uh validated go-to-market strategy on a very small niche, like an industry like the winetis, where me and my co-partners' expertise can be very useful to sell, to create trust and confidence, right? And we are trying to become ANSIS technology partners, for example, which again will help to get more channels and more traction and more sales. But once we have a traction that is well established, then we definitely want to go cross domain. So go thermal, go mechanical, and have our agents be the orchestrator of a much larger workflow where you have mechanical engineers sending jobs to the agents and passing them to the RF engineers that will send them back to the mechanical engineers for analysis after the RF systems are designed. We they will go back to other types of engineers. We want to make it easy and we want to make it, we don't want to dumb it down, we want to keep it at a high level so that every engineer feels like Ironman, so they can just send it out to their teeth's work, but they can teach the hardware.
SPEAKER_00Yeah, no, it's a brilliant approach. I say genius to everybody who agrees with me, but in this case, genius. In this case, I really this is something I really want to happen. Hurry up. I really appreciate you spending the time to explain RF to me and antenna design. I really love the fact that you're using engineers' trusted tools and helping make them easier to use. Starting off with something you know quite well.
SPEAKER_01I I learned a lot from my previous work. So if I can do anything on this podcast to thank them, Chris and Ridge Power and Hunter and Barun, the early team, we were just four guys in a garage, literally. I came as a mathematician designing algorithms and being very abstract, but little by little started doing embedded code, working on the RF systems, trying to help with the simulations and soldering the antennas that were blowing up. I learned just by being close to them. And I really want to create something that will benefit a company like that, many companies like that out there in the world.
SPEAKER_00Thank you for coming on. I know you're gonna do great things, and I want to keep up with what your progress. Thanks again. And have a nice resting. Okay, thank you so much. Thank you for listening to Faux Des the Future of Design and Engineering Software Show, brought to you by Ench Technica. I hope you have learned of a new application or technology that will help you with your job. If you have an application you think would be of interest to other engineers, please let me know by emailing me at repinder at enchtechnica.com or message me on LinkedIn.