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
Peggy Xia, CEO of gNucleus: We're Not a Text-to-CAD Company
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
A text prompt that “almost” makes a bracket is the fastest way to lose an engineer’s trust. I sit down with Peggy Xia, co-founder and CEO of gNucleus, to find out what, besides almost making a bracket, gNucleus was up to.
Peggy's background is not that of a typical AI startup. She has had a career in the real mechanics of engineering AI, CAD automation and simulation workflows with Siemens. He wrote the code for Solid Edge's Synchronous Technology. At Google, she developed YouTube's most widely used recommendation model, which emphasizes video quality.
We unpack why text-to-CAD goes wrong in ways that feel unforgivable to practitioners: the model is forced to guess when it has not seen enough domain data, and that guessing manifests as hallucinations such as misplaced fillets or incorrect features. Peggy explains the training stack in practical terms: pretraining, post-training with reinforcement-style scoring, and fine-tuning on customer-specific CAD data, so a model can become genuinely strong in a narrow domain such as automotive motors or assemblies. We also talk about multimodal AI, what “sketch to CAD” could look like, and why converting meshes or photogrammetry outputs into clean, manufacturable parametric models is still one of the hardest problems in the pipeline.
From generative design to computational design, we challenge the idea that a cool-looking shape equals an engineering solution. Accuracy, tolerances, benchmarking, token cost, and runtime matter, especially when you want production-ready results, not just visuals.
If you care about the future of CAD, CAE, and manufacturing-grade AI, this conversation will sharpen your mental model of what’s possible now and what still needs breakthroughs.
Welcome And Guest Introduction
RoopinderHello and welcome to FoDES, the Future of Design and Engineering Software Podcast. My name is Roopinder Tara. On the show, we will have guests that will discuss tools and technology that engineers will find interesting and useful.
Peggy XiaHello. Hi, Peggy. Hi, nice to meet you.
RoopinderThanks for uh coming on the show. First of all, let Peggy, why don't you uh introduce yourself and tell me a little bit more about gNucleus?
Peggy XiaUm I'm Peggy Xia, I'm uh founder and CEO of uh gNucleus. So um we're building uh engineering uh AI models for yeah, cat design, simulation, and uh future for manufacturing, discover and everything for the industry and the engineering vertical. Um my personal uh background is kind of uh mix for the AI and engineering. So in my earlier my career, I was building the uh CAD um software at Siemens, which is the next, yeah. I was on the CAD core modeling team, working on core functions like Extrude and Revolve, and later I was building the uh synchronous modeling. Yeah, I was the first uh person, I think, on Siemens working on developing the uh synchronous modeling. Yes. In other, uh that's a big innovation at that time for CAD. Yeah, it's also called uh direct modeling, so it gives you uh yeah, flexible flexibility to uh modify the uh yeah, like uh model directly. And later I joined uh uh Hexagon on their pre-processing tool, yeah. Uh, and
From CAD Kernels To Industry
Peggy Xiaalso the measure and uh generative design. Yeah. I was also on uh General Motors for one year to see how yeah, the PIM and uh CAD working on the big automotive uh like uh General Motors.
RoopinderUm so you went from developing CAD and CAE tools to actually working at General Motors?
Peggy XiaUm yeah, I'm first starting working on CAD and I switched to General Motors work on set on the enterprise uh gener Gem uh account for about two years, and later I switched to uh uh CE.
RoopinderLet's go a little further back. Where are you where did you study and where did you where where were you born and raised?
Peggy XiaOh, I was born and raised uh in uh in China. Yeah. I was first working uh in uh Shanghai in Siemens, Shanghai Development Center, and also in uh General Motors, China, and later I switched to uh US working at uh uh Hexagon's headquarters in Southern California. Yeah. And uh I remember it was at the year the 18, suddenly, you know, machine learning starts to become very popular. So I I was uh fascinated by the as that time still deep learning. Yeah. I say, oh, this is so cool, this is what I want to do. So I switched my uh job again. I first joined Amazon working on machine learning, but it's more on the non-personalized model. That is also the year that you uh you know being published. That's the base of the nowadays all the um yeah, models. Yeah. After at that time I remember I was still at Amazon. After I read that paper, then I said I want to join Google to join to work on large language models and transformers. So I uh switched again. I, you know, join at the same year I joined Google. Yeah. Yeah. I first uh joined the uh YouTube um recommendation team. I think I stayed there for the uh last uh six years working on uh YouTube uh recommendation
Amazon To Google And Transformers
Peggy Xiamodels. So basically only several models for the uh YouTube home ranking, yeah, home ranking models. And uh we are I'm very proud at that time that uh there are around two billion people every day using the model I build. Yeah, yeah.
RoopinderSo you built the model that uh has that recommends my next video. When I watch your video, uh it tells me you might also be interested. That's your work.
Peggy XiaYeah, that's my work. Yeah, and we try to recommend the high quality videos, yeah, on top of the low quality videos. That's what I'm my major. My my model is majorly identify the uh you know high-quality video like your videos to user on on top of the more, yeah. Uh oh, I see.
RoopinderSo you're probably not allowed to say it. I know Google likes to keep a lot of things secret, but uh how do they determine what high quality is?
Peggy XiaOh, we build models to determine, yeah. Of course, it's models. So everything is is it visual quality or is it uh uh a lot of signals, uh yeah, a lot of signals, yeah.
RoopinderOh, okay, yeah, okay. Yeah also the number of people refer referring to it or uh not exactly.
Peggy XiaThere are multiple models. We use um yeah, different signals, but yeah, I cannot expose more, but it's uh we definitely uh yeah, your model quality is definitely one of the important attributes uh for YouTube to uh recommend. I see, I see. Yeah. And uh yeah, also at YouTube at that time we work very closely with YouTube brain team. Yeah, at that time I remember in year 2022, yeah, internally we uh saw the first demo of Gemini. Yeah, yeah. So I remember uh the first time when I saw the early version of Gemini, I was so surprised because as a person working on machine learning for many years, if you want all the models are built for specific tasks, right? For example, for recommendation. So we train with how the model, yeah, specialize the task. It can only recommend videos, cannot do any other things. But I think Gemini is the first uh model, I saw it can do multiple tasks. Yeah. It can talk to you, can do math, and uh, a lot of the tasks it does not perform as uh uh well, but if
How YouTube Ranks Video Quality
Peggy Xiayou get more prompt, yeah, then it performs much better. So that's something uh we never saw before. So yeah, um so then we start to working on AI project inside the YouTube and one of the uh, you know, a few people started uh working on those for several years before even before Gemini is open to everyone internally. Yeah. So when ChatGPT is came out where everybody at Google is so excited, there is we saw the potential that this technology can be used outside, yeah, by everyone directly. Yeah. Before that is just the internal use. So that you know make me think a lot in the year 2023. Um thinking I need to do something more innovative videos to others using large language model. Yeah. So it uh I'm then I'm wondering what can we do? What can I do? So uh, you know, I recall all the pain points I saw when I was in the CAD and you know, uh CAX industry. You know, there are a lot of manual work, uh back and forth, yeah. And uh I think at that time still the large language model arm not to have the full capability to do this, but I think I saw the potential. Yeah, I saw the big potential there. So I so I make another big decision. I just quit my Google job and uh start uh gNucleus. Yeah.
RoopinderUh very impressive background, I especially what you uh not just the Google but also the uh all the Siemens and the NX. It's rare that we uh we get to talk to people that have that much CAD experience. And you said that that uh made you aware of all the pain points. Uh so NX uses parasolid geometry geometry kernel. Kernel, yeah. Right? And did you find that to be restrictive? Or because it can only model straight shapes the solid, it can't handle organic shapes real well. It doesn't do meshes. Or I think it now it does do meshes, but uh I I think that's more or less an afterthought. Uh what what do you think of the parasolid engine? Are you using the Parasolid engine in your current gNucleus.
Peggy XiaYeah, so there are different uh layers. So beneath CAD, there are geometry kernels and also constraint solvers like uh D-Cube and also another bunch of different stuff. Yeah. And on top of those fundamental components that build the CAD. And CAD is a much rich layer. Yeah. It will store all the history and all the operations, all the inputs and outputs, and then mapping back to the BREP phase. This is a CAD layer. Yeah. Let's put things in to two separate uh layers for SolidWorks, let's see. Even though yeah, but it can also use other uh geometry kernels. Like uh I don't see there's any limitations that it can use implicit modeling. Yeah. So I remember when I was working at Siemens, I'm working on another project called uh an Human. It's just you call create a human real. Yeah, the it's like a digital train of a real human in to do to help do all the simulations. That model is purely based on mesh. We create a human feature. So everything in uh in the uh, you know, it's a feature. We create a human feature, but it has all the mesh or yeah, the fancies. There's no BREP model underlying. So from the CAD side, I see it can support both. Yeah. But for the geometry kernel-wise, I think it's I think it's good to have one kernel focus on BREP like a Parasolid, and uh another kernel focus on other stuff. Yeah.
RoopinderOkay, okay. So Peggy, I have to tell everyone about a little bit of a small experiment I did with and what I did. This was a couple of weeks ago. Uh I've been hearing a lot about text to CAD, right? There's a lot of companies that are trying to do text-to-cad. You know Michael Finocchiaro, and by the way, he speaks very well of you. He's very he says he def he defends you after I did my little experiment.
Peggy XiaHe said, uh Your experiment is fine.
RoopinderSo I had almost on a dare, I said, okay, make me a bracket. And I gave it a very simple prompt. Make me a sheet metal bracket, uh, and it's gonna be this thick, and it's gonna be have this many holes. And I have
Gemini Moment and Founding gNucleus
Roopinderto say, it didn't do the bracket I expected, right? It put the round on the wrong, on the wrong part, it put a round where the fillet should have been. As an engineer, if I'm an I'm just an engineer, I write about this, these tools, but if I was just an engineer looking for a bracket, I would have at that point dismissed text to CAD because look, it can't even handle the simple bracket, right? So, based on prompts and feedback I got from that little experiment, including Michael's, I said, Well, and yours, you responded to thank you for responding. I went back and gave it another shot, and it was fine. It made the proper bracket. I got I'll show it to you if I can. Let me just show it to you.
Peggy XiaYeah. You know the difference? Why the previous one it cannot, but the later ones it could. Yeah.
RoopinderI think you did, I think you did a little more training on it, right?
Peggy XiaYeah, so that's the key uh you know difference compared to software. You know, software, you might see, oh Peggy, I find software has a bag. And then we have engineer uh fix that bag. But um model is totally different. So model is if the model didn't see such data enough, so it will do uh guessing. Yeah. So we normally call it hallucinations. Yeah. So um so the thing uh to improve the model is very simple and straightforward, just feed more data. So this is a very, I think, uh good example how data can improve immediately on this. So far, you know, the model we put on our homepage is a very simple baseline model. Yeah, we actually do a lot of fine-tuning for our uh enterprise customers with their own data. So it that will perform much better on their own data. Nowadays, I think you can see a lot of text to cat companies are doing demos on some uh, you know, either this part or that part. But we are, to be honest, we are not um uh text to cat company. We are a model company. What I'm doing here is fundamentally no difference compared to the you know recommendation system. We just get enough data and train the model. It's just uh the model we train is like uh CAD on the CAD domain. Yeah. And uh non-personalized compared to uh YouTube
Geometry Kernels Versus CAD Layers
Peggy Xiahome recommendations or yeah, it's home. YouTube is like personalized data. So we are non-personalized cat data. If you feed more data, then you perform better. Yeah.
RoopinderOkay. So gNucleus is not strictly uh text-to-CAD. It does, it makes more models. It could make like if I, for example, if I asked it to make me a uh motor, a big uh motor that's like one and a half times more powerful, generates one and a half times more torque than this motor, and I give it a picture. Will it be able to do something like that?
Peggy XiaWe internally can do can generate 95% of the e-party motor already. Yeah. There are several complex parts like housing we cannot do right now. Yeah. The housing, to be honest, I will call it uh one of the AGI problem for the uh CAD, you know. Um but for remaining of that we call it Autonova. It's the specialized model for automotive. Yeah, only for automotive customers. So you uh so this uh yeah, that one you cannot access directly on our website. Um but uh uh that model, because it generates uh assemblies, it's more complex. So far we can generate uh ePal model uh within two hours. Yeah. If you ask a human engineer to manually draw this, it takes, I think, uh weeks. But our motor model can only draw motors very well. Yeah, it cannot draw other things like robotics very well. But if we feed more data, then yes, yeah. And uh not only CAD, we are also working on uh simulations. We have we'll have a new version and come out in uh the next month with you know you create the CAD and uh uh run the simulation, and the simulation results give you uh um recommendations where you you can improve. So basically, we are more um general, I would say, CAD or simulation task model.
RoopinderI don't know that much about AI except what I read, and I'm an engineer. I have certain expectations of AI based on all the things that I hear. So that's what led me to text a CAD on all the companies that are doing it. But tell me about the training. You said you have to just train it. Is it like, okay, forgive me for this? Is it if I am training a dog how to do something, I keep giving it examples and I give it treats if it does well, and I, you know, I I give it, I don't spank my dog, but let's just pretend that I do, right? I spank it when it doesn't do well. Like the other day, my dog peed on the floor, right? So I didn't spank it, but I, you know, I gave it some harsh words. How is trading different? How do you do you just give it more examples and give it feedback? Or how do you train it to make a better bracket or a better motor or what have you? How do you do that? What are the examples that you give it?
Peggy XiaYeah, so there are different
Text To CAD Bracket Lessons
Peggy Xiauh types of trainings. For example, you might heard about uh so the model generally uh starts with a pre-trained. So it's non-supervised uh, you know, training. Basically, it means uh it does not uh have a label, it just trains with a bunch of uh sequence. That's the pre-trained model. Yeah, so all the big models that start with uh something like that. And after the pre-trained is done, then you will train it on a specific task. That's what we call post-train. For example, yeah, the one you mentioned about you want your dog to work on a specific task. If you're doing well, you give uh him some good stuff. He's not you give him some punishment. That's exactly what we are doing for reinforcement and learning. So if model is doing good, we give a high score. If model does not doing good, we give uh low score, and uh the the model get all the trajectory and uh get average on on this and it knows which way it should improve. This is called uh post-training. And uh after the post-training, uh normally uh the you know the general models like uh Gemini, uh it will, you know, uh ready for publish. Yeah. But on all specific cases, we are doing fine-tuning on top of those models, on specific tasks to make it perform even better. Because when all the models are trained, for example, the model trained, they won't have those specific motor data, right? They are only by each of the motor companies or automotive companies, those companies want to share their own proprietary data into any AI company. That's their own uh property. Yeah. So uh, but uh AI, if they never see such data, yeah, like the bracket case you mentioned, right? They never see the very detail of the motor data, how it how could it be performed well on that specific task. So then uh what we're doing is we get um customers' data and fine-tune on a model, then that model only will be used by that customer. Only they have access, then that model performs extremely well on motor uh generation.
RoopinderAll of the model creation is based on natural language prompting, correct?
Peggy XiaUh not exactly. Not as most of the models, they are uh much complex than uh it looks like. For example, uh all the models currently are multimodal. There are different ways you train Motemodel, you can convert all those things into uh yeah, use a tokenizer. So all the things uh fundamentally the Moodle C is just uh token or certain I could give it a prompt or I could give it a picture. Yeah, they will convert that to certain things only the model can understand universal uh space. Yeah.
RoopinderSo you can ask let me ask you something. So for a long time, I wanted to give CAD a front top and right view of a picture and say make this into a 3D model, right? So it could be a hand sketch. That's been a I don't know, for 30 years I've been wanting CAD to do that. Say, here's the here's a quick sketch of my part, right? And maybe it's not front top and right view, maybe it's an isometric view, right? So it's give it a quick sketch, show it to show you know, point my point my camera at it and say, make me a CAD model. How close are we to that happening?
Peggy XiaOh, very good question. Yeah. So you are a two makes. That happens, so we won't need. Let's see, let me do a quick estimation. If we only focus on the phone, and we collect
Pretraining Post Training Fine Tuning
Peggy Xiaall the phone data, and generate I will see the all the four pictures. I will make four pictures. One is front, back, top, or six pictures, and also an ISO view, maybe 10 pictures together with the details of the description of the of the uh the phone and then head model. Yeah. And uh I will collect a bunch of those data. And uh that's not also not enough. We will create uh a lot of small variants of this we call data augmentation and uh put that uh into the model to train. And then you will see something, and uh after you train the model, right? You need to know if the model performs good or not. Then we will prepare, yeah, eval set. Yeah, you can uh uh see that as uh you know a baseline. Then we will see the model's performance, they normally call it uh hill climbing. And uh so the model never sees the eval data, and you give more data, the model to train the model, and the model after model is trained, it uh uh running under eval data to improves. That's normally how we train the model. With enough data, I think uh we could do a lot of things. Yeah, like the case uh you mentioned about. So the then the first the question became how can we get those data efficiently? Yeah.
RoopinderOkay, right. You're you're familiar, I'm sure, with photogrammetry, because that was the promise of photogrammetry that I could take, I could give it not just three pictures or ten pictures, but I'd go around an object with my camera, multiple locations, spin it around 360 degrees at two different at two or three different angles, and it would generate a mesh model of my part, right? So I understand that, but that never really caught on. Do you think that's more possible now with with what with what you're doing? A few shots and then a parametric solid model being created from that?
Peggy XiaYou mean just get the mesh model?
RoopinderYeah, just from a mesh model or from a camera model. You think it's more likely now with AI than with photogrammetry?
Peggy XiaI would say uh this is still very hard. Yeah. The problem, um, let me maybe go back to the history of this. You know, when I was in Hexagon working on something called uh uh generative design. Yeah, a lot of people mix that that with, but it's a very old thing. From uh I think um, yeah, from a I think it this technology starts maybe 30, yeah, 30 years ago or even older. So basically the idea is you have a mesh model uh in simulation. And the simulation can tell you which part actually is not so important, then you can uh remove that, right? And then you get a more organic shape. And then how you map that organic shape first to surface model. And uh, if you have a surface model, then normally basically it's uh the uh nerve surface and convert that then to a parametric model, the solid model. So each step is extremely hard because it's uh reverse engineering. So before AI, I never there are a lot of tools doing that. Each company built some tools, uh, said it can do this, but still I don't see anyone can fully automate it that there's a lot of manual work you need to do. But when AI comes, how far are we from this? Yeah. So I would say still there's a lot of gaps. Yeah. The gaps uh lies too well. So for
Multimodal CAD From Images
Peggy Xiamachine learning, if it needs to learn something, the it needs to recognize a pattern. So this is it won't perform very well on uh mesh generation. Even though we have a lot of Lardys uh can do this quite well. But there are two things. First, there's no enough huge amount of uh data we have for 3D mesh models. Well, I see there are uh several good companies doing well on that, like the uh mesh, and uh yeah, it's one of the best. But still, this is for gaming or display. Yeah. The requirement for manufacturing is totally different. Manufacturing needs very accurate models. Uh like uh for the cell phone, yeah, we have customers which are smartphone companies, when we talked to them, they said if one millimeter difference on the model can make them, uh, the design team redesign the entire thing. It's not easy to for the diffusion model to accurately control the tolerance within one millimeter. So we see the big gap. You know, there's always a trade-off creativity versus accuracy.
RoopinderI tend to have two faces when I talk about CAD. One face I fall in love with the technology, and then I turn around and I'm very critical of it. So this happened with generative design. At first, I thought this is wonderful, you know, that this shape can form based on principled stresses, right? It puts the material where there's a high stress and it removes material where there's low stress. And I thought, this is perfect. I love this. But then my on the other side, I looked at the shapes that it created, and they were, like you said, organic shapes, they were arly, they were too, they had knobs and protrusions, and they were like and they were stringy. And I thought, well, that's never gonna be manufacturable, unless it's 3D printed. Yes, unless it means then it only does well at a certain with a certain load case, but if you change a load case from tension to compression, for example, all those thin members are gonna buckle or snap, right? It doesn't work in all load cases. So and then I fell in out of love with generative design, right? But then it occurs to me now. Bear with me for a second, it occurs to me now that here's uh our AI could step in here and say, if it's making a something that looks like a cable, right? It's starting to make something that looks like a cable. Cables are great intention, right? Then jump ahead. Don't do that last of your calculations, just assume that it's gonna be a string, right? Can't AI come in and determine, hey, that's a that's a cable or a string or something, or one another favorite example I have is my bicycle, right? I tell everybody my bicycle is a perfect design, and I challenge them to improve upon it. A diamond frame construction of a bicycle is perfect, right? I love my bike. And so I tell people, can you make a better bike than this? And nobody's taken up the challenge. And that to me is a very simple shape, right? So I'm I'm asking like two things here with one question, really. Like, why can't finish the job that generative design starts, right? Why can't AA look at a shape and say, oh, this is a this is a cable or this is a this should be a round tube, right? Why can't it just jump ahead and do that? Because to me that seems like then it would be not have knobby protrusions and things like that. It would be straighten itself out, or it would be become a beam because I like beams or you know, something like that. Why can't why can't it do that, finish that last step and make it what it's trying so hard to be?
Peggy XiaThe simple answer is because nowadays the AI never saw such data or trained on such data. Yeah, and so there are very few uh design data that AI saw. Certainly is uh certainly is you know you need to convert those tasks into something AI can understand. So basically, nowadays AI are trained on something we call the deterministic um task. That means when you ask, let's see, go back to a bike example. You if you want AI to improve on the bike to make a better design, you need to uh, let's see, have 10 different design and score it accurately or deterministically. That means no matter it's you or me, based on your score criteria, we can get the similar score or rank. So we always rank the this design as the horrible design or that design as a better design. Someone needs to convert those uh uh use cases into task.
Why Mesh To Parametric Is Hard
Peggy XiaAnd then AI needs to train with large amounts of those data, then it can perform well on that data. Yeah. So this is actually what uh we are currently doing because um we know you know when all the big models are trained, like uh, I think Gemini, GPT, and uh cloud, they are the same. It's when they train the model, they normally won't optimize for one specific vertical. Yeah. Okay. So there is a thing called data mix when they train all the data. So they basically need to decide how much percent of this data when they train the model. Yeah. So there's a good, make sure there's a good balance or trade-off on the model overall. That's good for the general purpose model training, but it's bad for specific uh vertical like engineering. Yeah. You need a very detailed specific task definition, but the general model won't be able to provide. Uh so this is the reason I think for uh engineering specifically, we need something dedicated for this vertical because of the uniqueness and the uh requirement for high standard, like uh very detailed, accurate work. Yeah. If we train with a lot of uh design data and we can uh successfully convert the uh score, the good task into the uh bad task, yeah. I would see maybe tens of tens of thousands of data, then you could do this, yeah, reliably.
RoopinderOkay, okay. Is it so okay, you have to train it on different things, different models, different assemblies, different products, even. How close are we to having being able to train AI on the products and shapes that I might have in my company, right? If I want to, I'll give you another example that I've posed, like an automotive company. If I'm at an automotive company and I'm I want to electrify a car model, right? That's something I'm my company has already done before, right? It's taken a car, but make this car into an electric car, right? It seems like all that history is already there in that company, right? That I know how electric cars are made, I know how internal combustion cars are made. Why, how close am I to having an AI, say this model, and it knows that okay, swap out the internal combustion engine with electric motor. I don't need the radiator, I don't need an automatic transmission, I don't need an exhaust system. You might ask me questions too as it goes along, right? Like how how much torque do you want, or how much my point is it's got all that history in my data, in my products, in my team center or whatever, right? It's got all that information and it's got both types of cars in its system.
Peggy XiaYeah, you so that's a special expertise. So yeah, you need uh you know, vertical AI companies like us under who understand uh how AI train how the data works for AI and also understand the industry, right? If you ask a general purpose one, you can compare uh convert everything easily into structured data like a JSON, and you throw that into a model to train you, then you get nothing. This is called a garbage in and garbage out. So even though your raw data is very high quality, but you have to convert it into some way the AI can understand, or can uh you can see the very clear performance gain on this. And we are yeah, we are this company are here to help, to help you to convert your data into something AI can understand and show you, create a data set for you and show you how the uh we made a baseline or benchmark you can see after you threw a certain amount of the data you the model performs are you know uh like this, not like this or like this, you know. Yeah, so how we yeah uh start to build uh gNucleus, yeah. And uh what we want to uh the gap we saw there and the gap we want to uh bridge. Yeah.
RoopinderOkay. Yeah. But
Generative Design And The Bike Test
RoopinderI I like your answer because it it does, it shows me that there's still a challenge to it instead of other companies. I have to say other companies, including one big CAD company, and I've talked, I've written about this in the past, so I don't mind saying it was Autodesk, right? Autodesk after after Chat GPT came out, you know how everybody was like went crazy. Obviously, you know about it. The world went crazy for and at the next Autodesk University, Autodesk came out with a make a car, make me a car AI. You actually ask it to make you a car, and it would start making the shape of a car, right? And that was their answer to ChatGPT. Looking at it, I thought, this is ridiculous. This is it's not a car, it's the shape of a car, and it's at best it's good for aerodynamic study, right? But that's it, there's nothing underneath, there's no motor, anything to electricity, there's nothing to it, right? That's but that was their answer. Like, hey, look, here's AI. We have AI, we can make a car, right?
Peggy XiaUh people it's ridiculous, right?
RoopinderAnd here they are a big important cat company saying, you know, hey, make me a car, and then next, and I expected them to say also next year to say, I could we can make buildings. We have we know how to make buildings. Hey, AI, make me a building, right? So there's a lot more to it. And I really appreciate that you think it's a challenge, right?
Peggy XiaEven with all the customer data, it's not it's a challenge, and there's one path, maybe in the future I see it can solve this problem. You know, there are a lot of uh right now the new labs now working on kind of like uh auto-research. Yeah. It's just searching the AI to learn how to solve the problem like itself. So this is the you know, kind of their definition of AGI currently is like AI can define the problem and explore the way to solve the problem by itself. And if that if we can train that model, then it might solve your problem. So basically, uh it will first uh study your data. Yeah, figure out a way to extract the data by themselves. Nowadays gNucleus can do this, but AI cannot extract the data for you. AI only can be fed the data like a baby food made gNucleus. But if the baby grows, baby knows how to uh get the data from your raw data and train a machine learning model. Or it might not be uh, you know, um basically it can be any model or any methodology and then build a system for you, then yeah, that's the AGI time. Yeah, it's coming. Yeah. And that methodology we see can fundamentally apply to any vertical. Yeah, not only CAD or simu, it can be applied to everywhere. You can see there are several uh Neolab, including the one I think uh uh Jeff Dean recently uh built, uh working on this erection. Yeah.
RoopinderOkay. You're very familiar with generative design. How different is that from computational design? You know, nTOP, the company nTOP used to be called nTopology by Brad Rothenberg, he's their CEO and founder. They say, we don't say precisely this, but it they I think they think that's AI, right? That computational design is is AI because it can make optimal shapes. I'll tell you where I'm going with this. I issued them the bike challenge too. I said, hey Brad, make me a better bike than I have, right? And it and he and Brad said, sure, Roopinder, I'll do it. Uh let me do it. Because he's I never heard from him again. He never came back with a better design. So I'm thinking that even though it's a parametric, it can solve the optimized parametrically. It never was able to make a better bicycle than I have. And so what are your thoughts on parametric optimization? Do you think that's AI at all or can be AI?
Peggy XiaUh, I think there are these are two actually different uh, you know, things. It's not uh apple to apple comparison. It's an apple to orange comparison, right? So basically AI are one of the methodology you can use to solve one problem. It's the the way AI can solve a problem is you give it a lot of data and it has a neural network. And then it can summarize certain patterns, then solve a problem. It can be a parametric one if you have enough model. So for the uh computational or implicit modeling, or I I would say it's uh more like an algorithm. Yeah. So if uh AI model, let's say an agentic model, it trains on a lot of uh, you know, knowledge and it's aware of how to use this tool properly. So that AI model might decide, oh, if I want to design a better bike, there are several paths I can do. So the first thing the AI, the good AI will do is go through the internet to get all the papers, to get first what is the definition of a good design. Secondly, for each of the good definitions and what are the approaches, it will list all the good approaches, then give you a summary or report. See, these are my study results, and you as a human being, you make a decision. Here, I give you recommendation. Let's see, let's use inflicted modeling, or we use uh uh reverse engineering, use a bunch of tools and give you, and you then take a look at the report and make a decision, then AI go do it. Yeah, that's the way I think yeah, a good AI can help you to design a better bicycle.
RoopinderI gave the same challenge to Autodesk. Make me a better bicycle. And Peggy, you should see what you might have seen what they I don't know if you follow the Autodesk very much, but you should, I do. But the designs they came up with were ridiculous designs. They started off with something that looked like a bicycle, but
Turning Enterprise CAD Data Into Training
Roopinderit was through generative design, and it started having so many weird shapes on it, right? All the time, you just start with round tubes. Uh, round tubes work great. Uh, start with the round tubular tubes uh for your frame members and just position them uh differently parametrically. But they did, they made this blobby-looking bicycle shape. And in the end, it was never, you know, they were all about light weighting. They were all about we could use gender design to cut your weight of any, cut the weight of any part, and they bragged about parts that were 30%, 40% less weight, right? But all the time the blobby shapes they were making for the bicycle were heavier and heavier. In the end, they just gave up. They have this design gallery. I think they just said it's art, at least it's art. It may not be really good optimum design, but we'll call it art. So it seems like their stuff didn't work because we already have optim, close to optimum. Optimal designs already, right? And it and it didn't work in optimizing them any further. So it was just technology gone bad. And I see a lot of potential in AI to, like I said, finish the job. Like AI starts gender design parametric optimization in in simulation when uh it converges to a solution asymptotically, right? AI can step in and save 90% of the computation just by going to where that curve is going, right? It's heading towards a solution.
Peggy XiaYeah.
RoopinderJust take it, do it.
Peggy XiaYeah, they call it uh currently the physics model. I thought there are many companies like uh I think uh Physics X to doing this, and there's also a data company is called uh Luminary. It's generated a lot of synthetic data for simulations. Uh but still I would say uh the data set I saw uh generated by Luminary is based on uh several narrow um, you know, the um for cars, and I would think it'll definitely perform very well on those specific cases, just like our motor case. But uh, if you want a general purpose one to work uh well, there will be more data. Yeah.
RoopinderYeah, yeah. Another example is uh you know how the uh AI is used for uh not ray tracing, but uh photorealistic rendering. You know, it I think Nvidia has a very good approach to this. They'll actually see where the picture is going, right? Like and finish the job. They will actually, you know, they don't do ray tracing can go on forever. The way the ray bounces, it can go on infinite times, right? But they'll say, oh, this is where the picture is heading, this is how it's gonna converge, and they'll sharpen the picture. They'll save 90% of the computation of ray tracing and finish the picture. So it's almost like it's like rendering in real time because it doesn't take all the computation that ray tracing takes. Yeah. That kind of thing. Like if it can be done for pictures, all the time I'm thinking, oh, why can't it be done for models? Why can't it be done for 3D shapes?
Peggy XiaYeah, for yeah, but you know, it's much easier to generate uh, you know, 2D image or rendering data samples than the 3Ds. For simulations, each of the data points you generally need to you need to run the simulation, right? The simulation, each simulation, the simple ones, R s and the complex ones, days, right? Each data uh sample is very expensive. Um this thing I don't worry that much because the trend is there. It's just a matter of time. And uh I would think uh you know, compared to two years ago, yeah, to nowadays, big improvement uh even on general purpose model, how good they are doing on CAD. Yeah, right. Previously they cannot generate any CAD. And now that's by the agentic AI by write code, they can generate a lot of things, even still not uh in production ready, but you can see the improvement.
RoopinderI just use my phone again, but there's a lot of examples. Like if I'm saying, okay, make me make me an iPhone type phone, and it starts meshing, at some point it's gonna say it's it should realize, hey, this is a totally flat surface. I think it's trying to get a totally flat surface, but it won't be. The mesh will be a little bit bumpy, right? At some point it could just to me, it seems like it shouldn't be that difficult for it to sense this is heading towards a completely flat surface. And my design rules say I want flat surfaces. So step, don't make it finish the mesh. Don't do 90% of the computation. The rest of it, just
Agentic AI And The AGI Bet
Roopinderassume your because of your design rules or your design history, that you want a flat surface right here, right? And over here, it's developing into a round. It's go ahead, make it finish it off. These are my design rules, right? That should be, I don't, it's uh excuse me for this. It seems to me that should be a very simple operation, right?
Peggy XiaYeah, it's a simple operation, yeah. But need to build it uh together into um yeah, let's see uh phone tools or phone tool sets, then the AI understands how to make it. And the most complex part, I think, uh, for phone design specifically is the bullying operation. If you have very small uh you know widgets and how to uh cut inside and put the tiny piece into that fits perfectly into phone. That's the most uh complex part. But now no AI company currently have that data. You have to, you know, let's see if I want to do this, I might, you know, uh, you know, first uh hire a bunch of phone experts and generate that in-house data by myself. Still, it will be very slow because those data are very, I would see, highly professional data, and then build a data set, yeah, and then train the model, and then do that. So my methodology is very dumb. You will keep here um saying, oh, lack of data, get enough data and train the model. You'll see the magic. Yeah.
RoopinderIt's yeah, it's so uh yeah. I apologize again for making it sounds seem so simple and demanding so much. I understand it's it's quite complex. I gotta say, you might be, but based on what you're saying, maybe you'll be the first to figure it out. Maybe gNucleus will be able to do that, right? I think you how many people do you have working on the problem? How many uh people in your organization?
Peggy XiaAnd oh, we currently are still uh a small company. We have about uh uh 10 to 20 people, yeah.
RoopinderAll in the States or in San Francisco, right? Your headquarters in San Francisco?
Peggy XiaUh we are in based in uh Sunnyvale, yeah.
RoopinderSunnyvale, okay.
Peggy XiaYeah. So half uh of our team are AI person and half are uh engineering. My co-founder Mei, she was uh my previous colleague at Google, she was more uh AI person than you. I'm still a cat and or simulation AI person. She's a pure AI person. She was working at uh uh Google Bring and uh Gemini for 10 years. Yeah, her work is in a lot of the Google uh product. Yeah. Uh um so basically we are very uh small group but work very uh efficiently. Uh everybody is uh expert on either AI or uh engineering. Yeah. Okay. So if there are people who want to uh join us, sorry for the advertisement. Yeah, yeah, feel free to uh go to our website and uh contact us. We really um the most important thing I think for us to find uh people is uh uh first it has a motivation or it um has uh enthusiasm belief this is a future direction. Yeah, and then it has um uh in enough knowledge is either on AI or yeah, industry software, then yeah, and uh the third important is can work together with us.
RoopinderOkay. Yeah, you're you're I think you're well funded for this stage of your operation. What's the next goal you have? Is it an end product or is it uh technology that's licensed, or what do you hope to make gNucleus into? Or is it a consulting service?
Peggy XiaUh we are a model company, yeah. We are actually a very quiet techno, high-tech, deep tech company companies. A lot of people think we are uh text to company,
Computational Design Is Not AI
Peggy Xiabut uh, we are not. We are a model company. So our next uh bigger mission is train uh is collect more data or generate uh more data and train a bigger model. And we are not uh, you know, uh other companies can also use our model or APIs. So we're not that interested in build text to cat or text to assembly. You saw our website just to showcase how the people can use our model. It's just like uh chat GPT versus open AI, but open I is not uh uh chatbot company, it's uh a model company similar to us. We are engineering AI model company.
RoopinderSo you want gNucleus to be like the chat GPT of models.
Peggy XiaSo yeah, we are uh like open AI for engineering. Yeah, like open AI, yeah. We build models and uh any other startups or CAD companies or your companies, if you are interested in using our tool, you can uh using our models and you can, you know, um you can just use our uh APIs. Uh yeah, that's our um goal. You know, there are always different layers inside AI word is the same. Most m I would see 60% of the companies working on application nowadays called agents. Yeah, it's application, to be honest. And beneath that, I would see it's a model company and the infra company. Yeah, we are more like the model and infra company. We build our own infra. We can deploy our entire system into enterprise, their private cloud, or even on-prime. So it can um build up pipelines to automatically extract um you know data from their existing data, like your case, right? You have a lot of data, but you don't know how to convert that into something the baby AI can eat. So we are here to deploy the entire system into yeah, enterprise and uh extract food, uh data, convert to something AI to understand, train the model and uh then deploy the model. Yeah. Then on top of that, build the agent. Yeah. We compare with the public model actually uh on some of the our own in-house preparatory data. So compared to uh public models, we use a smaller model, which can um get uh 30% of the success for rate, yeah, compared to the public model on certain narrow uh slides and uh uh much faster. Nowadays, most of the coding agents they can do some things very well, is because they keep retrying it, they call it uh harness, right? Harness is basically you keep retrying it, you remember your past failures, you avoid your past failures, and then you but they normally in order to reach a success result, you normally go to 10 or even more round, yeah, agent loop. But uh with our trained data, we can train the model, we can largely reduce this maybe into one or two loops. Yeah.
RoopinderOh, okay.
Peggy XiaSo that's much cheaper, right? If you can do 10 loops, 30 minutes to generate uh uh catapult versus we two loops, five minutes. Yeah. If you take a look at our CAD Bench, our website, right? We publish a lot of bench. We don't publish our own model performance. So that's normally our criteria as a bench owner. You don't put your model on the CAD Bench. Otherwise, no people really trust your bench. You just put other models. You can take a look at the time and uh uh the uh cost and uh also the parts it creates for a gear, right? A simple gear. Normally the frontier models now it's about uh 15 to 30 minutes. Yeah, yeah.
RoopinderYou talk about a gear like a uh spur gear or sprocket.
Peggy XiaYeah, okay. Yeah, we published all the data on Hugging Face. You can take a look at the time and cost and how many tokens. Yeah, nowadays they can draw this uh very well. The recently our bench shows the Opus 5 can complete 80% of the uh 90% of the case. Yeah. But the cost is more than $100.
Benchmarks Cost Tokens And Deployment
Peggy XiaYeah. To draw, uh we have 10 cases, very simple, like uh knots and bolts and uh um uh gears. Yeah, each each part takes I think 15 to 30 minutes to generate. Yeah, that's a non-fine-tune model. But if we fine-tune this will be uh very quick, five minutes, less than five minutes, two minutes. Yeah.
RoopinderOh, gNucleus, it takes five to ten minutes.
Peggy XiaYeah two to five minutes. Two to five minutes.
RoopinderYeah, yeah.
Peggy XiaWe didn't put that uh on our yeah, leaderboard, but yeah, but internally we we have this data. Yeah.
RoopinderOh, okay. That's that helps a lot. I'm looking at the list of of all the models that you tested, and gNucleus isn't there, but you use internally your data.
Peggy XiaYeah, we have internal data. Yeah, we have our own internal CAD-specific uh agent as well. But normally, you know, when you do a leaderboard, you normally don't put your own model on uh on that leaderboard, you know. That's a rule for a leaderboard, yeah. Oh, I see, I see. Yeah. And how why we make this bench is because internally we fun tune on different models for different customers. So we internally have this data. So we I wonder how about we put this publicly so to save people a lot of money, you know, in order to make a bench, you um you will spend uh, let's see, uh for our simple bench, it's about uh over uh one uh $1,000. We have put all the monies, uh all the models and combinations. Each is about 100. We have 50 models. So this can uh we publish this data. This could save uh a lot of people's money to rewrite. Yeah.
RoopinderNobody realizes the real cost of that. I would but you know, but I do. I think I think like, oh my god, how many how many trees am I burning to make this bracket? Yeah, yeah. Yeah. Also, and eventually I'll be thinking, if I were a paying customer, I'd think how many tokens am I using? But I also think about the environmental cost of these engines, right? It's ridiculous. Uh nobody understands that. Nobody is talking about the real cost of this because it's hidden in subscription.
Peggy XiaYeah, it's hidden, yeah.
RoopinderIt's hidden.
Peggy XiaAnd and if you take a look at our little bottle one, another thing interesting is even though Opus 5 gets the highest score, the second one is uh Grok 4.6, right? It's a slightly lower score. It's about 87 score, but the cost is only uh, let me take a look at my account. If I remember correctly, it's about only 30, yeah, only uh it saves about 30% of the cost, right? If you are really using this in production, you don't care about it. That's slightly worse on Opus 5, but it saves you 30% of the cost. You were no-brainer, I would choose Grok for uh point six.
RoopinderSo instead of LLMs, which are large language models, you're is it accurate to say you're making an LGM, a large geometry model? Is that the fair?
Peggy XiaGood question. So actually, our model is not a language model. Nowadays, all the models are like multimodal models. So we normally call it AI model, JAI model. Yeah. So for our specific case, what we're doing is we want to build uh engineering JAI model. Yeah.
RoopinderOkay, all right, very good. Well, Peggy, that's that's great. I I really appreciate your time and thank you for explaining
Closing And Listener Outreach
Roopinderthat. Uh, I expect great things from you.
Peggy XiaI'm really enjoying the conversation today. Yeah.
RoopinderOh, so thank you very much. Thank you. Peggy, great, great meeting you, great seeing you. And I hope we stay in touch. Bye-bye.
Peggy XiaYeah, stay in touch, but bye.
RoopinderThank you for listening to FoDES, the Future of Design and Engineering Software show, brought to you by ENGtechnica. 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 roopinder at engtechnica.com or message me on LinkedIn.