Viral Shah on Dyad - Physical AI for Systems Analysis

“Make me a car” is an impressive demo until you ask where the braking hydraulics, controls, and safety logic went. 

We let Viral Shah, CEO and founder of JuliaHub, and Chris Rackauckas, tell us about physical AI and its use for systems analysis. 

Its a different AI story than the popular one: physical AI for engineers, where models must respect governing equations, compile, and validate against known test cases. Along the way we unpack why Julia, a programming language, was created and how it led to Dyad. Hint: to do systems analyses.  Any system. Also how open source shaped its growth, and why that foundation matters when you want AI to do more than autocomplete code.

We then get concrete with Dyad, JuliaHub’s domain-specific language for systems modeling and multiphysics simulation. Rather than building another CAD tool, Dyad focuses on function over form, the system and subsystem level where real products live. That unlocks fast iteration in the engineering V-model: requirements, architecture, integration, and ultimately digital twin workflows, without forcing every engineer to become a full-time programmer.

The highlight is a demo where an agent ingests NASA HL-20 lifting body documents and aerodynamic data, generates a working systems model, runs a documented pitch-pulse test, and produces plots you can compare to the original validation figures. We also talk about the trust problem with AI and why physics-aware compilers, transparent artifacts, and test cases change the conversation from “wow” to “verify.” If you care about engineering simulation, systems engineering, agentic AI, and digital twins, subscribe, share this with a colleague, and leave a review with the tool you want AI to tackle next.

Transcript

Welcome And Show Purpose

Roopinder 0:07

Hello 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. Viral is the CEO and founder of Julia Hub, and he has recently received $65 million in funding for a project called Dyad. Dyad promises to bring physical AI to engineers. Hi. Chris, I recognize you. I was watching your uh video. You're attempting to teach me about uh Julia Hub.

Chris Rackauckas 0:49

I have a few a few of those. Which video?

Roopinder 0:51

You're trying to show how you could use Julia Hub and I think Dyad AI in uh it was about 13 minutes long. Gotta admit though, it went over my head because I'm not a developer.

Chris Rackauckas 1:04

Okay. Yeah, it's also a pretty deep uh mathematical topic, but we'll we'll keep it higher level here.

Roopinder 1:10

Good to meet you. Dimitri, thanks for being on the show. I'm Indian, so we can talk. Yes. So you came to India, you came to California, you learned how to play Ultimate Frisbee, right? Yeah, yeah.

Viral Shah 1:27

If you're in California, as Chris, who's a native Californian, will tell you gotta surf and play Ultimate Frisbee. Those two things, you know.

Chris Rackauckas 1:34

That's it. Yeah.

Viral Shah 1:35

Otherwise, you know, you can't get your uh California, you know, lived in California stamp, you know. You cannot.

Chris Rackauckas 1:42

Oh, you also have to eat at In N Out. That's a requirement.

Viral Shah 1:44

Yeah, you have to eat at In N Out. That's right, that's right. That's the three things.

Roopinder 1:48

And sushi. Sushi was the California's first place ever at sushi, so you gotta do that. What what what year did you come to California? You went to UCSB, right?

Viral Shah 1:59

And then yeah, I did Santa Barbara 2001 through 2007. Oh I went there for my PhD basically. Okay, all right, yeah, yeah.

Roopinder 2:08

All right, one thing you

Meeting Dimitri And Origin Stories

Roopinder 2:10

you may not want me to highlight this from this article, but you're the first Indian I've ever spoken to that admitted that he got bad grades in school.

Viral Shah 2:20

Yeah, that's uh that's a bad Indian, right?

Roopinder 2:23

Yeah, well, I well, you certainly have made made good on things, so I imagine your parents are once again proud of you, right? So, but that must have been a a bad time for them where you're getting those poor marks, right?

Viral Shah 2:36

I I was a star student until high school when you know, when they were paying attention to me, then they let me be by myself and I let my better instincts take over or worse ones, I don't know. And so I just did not pay attention to school and just kind of had fun doing open source software and it put me on the trajectory I am, but it, you know, it there were quite a few close shaves as a result of that. But I I tend to mention that because a lot of people like you know have so much pressure growing up in India about like grades and like you know, like kind of going through the beaten path, and it doesn't have to be, right? Like, and and at least for me it worked out doesn't mean bad grades will always be a good good way to get into something interesting, but it doesn't have to be as bad as it sounds.

Roopinder 3:19

You didn't do well at uh digital signal processing. I I didn't I didn't do well at differential equations either.

Viral Shah 3:26

I mean oh no, oh no we you have you have the world's best differential equations person on this call. That's Chris.

Roopinder 3:32

Oh really?

Viral Shah 3:32

Okay, yeah.

Roopinder 3:34

Okay, this is why I can understand what Chris was saying.

Chris Rackauckas 3:38

Everything I do is differential equations, yeah.

Roopinder 3:40

Oh yeah. That that's his natural language, huh?

Chris Rackauckas 3:49

Yes, yeah.

Roopinder 3:51

So you went way over my head. I gotta say, I gotta watch that again to try to get even a glimmer of of hope. But uh his I watched his video. I was trying to learn what Julia was all about and and diet was about, and uh I had to slow it down.

Viral Shah 4:08

Yeah, very cool, very cool. I'm glad I'm glad you put in the time.

Roopinder 4:11

I imagine you're getting quite a bit of attention. Uh there's nothing there that demands attention and uh like getting a 65 million dollar funding operation, right? I imagine everyone's everyone's after you now.

Chris Rackauckas 4:23

Yeah, yeah. You got questions, yep. Yeah, everyone wants to talk.

Roopinder 4:27

Just one more thing, and I'll get back to business. So you actually married married a nice Indian woman. Family has you have two PhDs in the house now. She's also a doctorate. That's right, that's right. Yeah, okay.

Viral Shah 4:41

She's a microbiologist. Ah, okay. Um and and she's she's actually a scientific editor. And weirdly enough, like Chris's first paper was published by her. Oh, is that right? Yeah, yeah. Yeah, like we got you gotta go full circle.

Roopinder 4:58

I'm I'm trying to get

India’s Tech Gap And Open Source Culture

Roopinder 5:00

off the subject and I can't because I read also in your bio you have a book called Rebooting India. I hear these stories about China, how it's you know, re it's beating us on autonomous vehicles, it's it's get it's getting close to us in AI. It, you know, it's got that deep seek engine, and it's yeah, and I'm thinking, but where why don't I hear more about India? Why don't I hear more about India doing stuff with AI, right? And and there's a big technological gap. You allude to this, you actually reference this in your book. Like, we are a nation that has so much of potential and yet has made such little gains in tech, other than supplying the world with it's with people that work in IT, right?

Viral Shah 5:47

Yeah, other than that, right? Which may not which may no longer be necessary given the way software agents are going.

Roopinder 5:53

Okay. Right, okay, right.

Viral Shah 5:55

But yeah, no, you're absolutely right. But like I will not quote the person who said this, but I'll tell you uh a quote that I often heard when I used to be in India around uh around open source software. Someone said this we are a nation of downloaders, never gonna upload anything back, no, never gonna write it, just take it all for free. And uh um and right, like I mean, like India tops all the download charts for all kinds of open source software and everything. But how often do you hear of about a large open source project which was truly based out of India? And you know, I mean, I I built it out, I built Julie out of India for like almost like seven, eight years before I moved to the US, and uh and we created a very vibrant community in India. If you you know, um in fact, uh Chris and I also started the scientific machine learning organization where a large number of the recruits come out of uh the the ITs, not not the fancy ITs, but like you know, the all the new ones, right? Like and uh and and even sort of some of the smaller colleges that that you've never heard of. And we get some tremendous contributors out of there, and so we feel like we've we've really like brought India into a world of open source, at least in our world in a way that no one else.

Chris Rackauckas 7:11

It must be 75% India, yeah. Indian, yeah, in our in our GSOCs and in our contributor list, yeah, yeah.

Viral Shah 7:16

Really? Yeah, so yeah, it's it is off the charts, the kinds of numbers. Uh you won't see those numbers in other open source projects.

Roopinder 7:24

It's a democracy. The US is a democracy too, right? But there's several things that go against uh entrepreneurship and venture capital.

Viral Shah 7:33

You know, it always boils down to culture, right? Like, you know, once you create a culture of creating open source software somewhere, like other people get connected to it and they start doing stuff, and then it becomes a thing, right? And you know, you get a couple of colleges where it becomes a big deal, and then everyone's doing it, right? And uh it's it's just a matter of culture training, mentorship, like all these things, which are hard to get. And and that's why students out of Indian colleges are so keen to get mentorship as part of the Julia Julia Summer Court, the Cyamel Samurai Code that we do. Um they're just really thirsty for like for getting that high-quality mentorship.

Roopinder 8:11

Yeah, but it's almost like it's a country over with itself because as much as it wants to have wealthy people, it can't have people that are too wealthy or too well off because there's so many people that aren't. So then the people that aren't will rise up or rebel or revolt against uh you know the inequality, the injustice, right?

Viral Shah 8:32

That I have different, slightly different views on these things, as much as you know that the richest in India live like the richest in the US and the poorest in India look live like the poorest in Africa, but the masses are somewhere in the middle. But the thing is that like Indians do not like as a I mean, and I'm sure like I I assume you grew up in India, right?

Roopinder 8:51

I tend up to a young age. I was all I was seven when I left India.

Viral Shah 8:56

Oh, so you yeah. I I I don't know if you've sort of you know noticed this about Indian culture that it's it's never about wealth, right? Like like happiness is not tied to wealth uh in for most of Indians. It's uh you know, it's just money is just not as important a you know thing for people in India as it is elsewhere in the world. And people are happy, you know, they have happier family lives, they have stable lives. There's uh there's many good things and there's many bad things, but I don't think uh like people measure success with money as much in India as they do elsewhere.

Roopinder 9:32

Right. My mom was happy when I became an engineer. You gotta be an engineer or a doctor, right? So or or computer scientist, I think nowadays they nowadays that's acceptable. That's acceptable, right? But yeah, she was uh she never actually asked me how much I made. You're you may be right. I and I never told her, right? But as she just was happy I was doing well, and I got a good job, and I had you know a nice wife, right?

Viral Shah 10:00

Right? So that's kind of like how like it's a different sort of cultural, you know, approach to life, and uh you know it is kind of rooted in the Indian culture and the Indian way of doing things and our stories that we tell ourselves, and you know, the the stories we surround ourselves with there, right? So I just feel it's rich and it's it's it's not right or wrong, it's just different, is kind of how I I think about it. Yeah, now I I'm a Gujarati, by the way, so doctor and engineer not acceptable answers. If you don't run your own business, you're a nobody. I'm sure you know that. Ah, yes, Marvadi families, right?

Roopinder 10:33

Like you guys are business, businessmen, business people, right? Right?

Viral Shah 10:40

You're better off, you know, having a corner shop, uh, you know, rather than being the CEO of a multi-million dollar corporation.

Roopinder 10:48

Uh, we're Punjabi. My dad went after Harvard and got his Harvard PhD. And when he came back, he was he was top guy in the village for sure. No question, right?

Viral Shah 10:59

Oh wow, that must have been impressive.

Roopinder 11:01

I mean, oh yeah, it was literally a village. So I think they did the I wasn't there, of course, at the time.

Viral Shah 11:07

That must have been like the whole village must have come out to celebrate when he got back. I think I think they're hoisting him on his shoulders. Yeah, yeah, I'm sure that's actually like

Why Julia Had To Exist

Viral Shah 11:15

you know, with the garlands and like the the full the full I gotta get back on topic because I don't have you all day.

Roopinder 11:21

And yeah, and and I want to learn more about Julia, the programming language. You created a whole programming language because because why? There were already so many, right? You're Fortran, basic, you had S, you learned C, you you uh it probably C, right? Uh there's Python. Why did we the world need a whole programming language?

Viral Shah 11:45

You know, it's uh you know, I I call myself an accidental programming language designer because it's not what I you know fancy myself as doing, right? Like I, you know, like the others who started Julia with me, like Jeff Bezance and Stefan Karpinski, like they are professional programming language designers if there is such a profession. Right? But like they they they get out of bed thinking about the next big programming language. And you know, I've you know, I'm more like Chris. Like, I you know, we think about applications, like we want to make scientific computing easy for scientists and engineers. And we looked around, like in my PhD, I did a lot of MATLAB, and my PhD was around parallelizing MATLAB itself. And I mean, I was really just fighting it all along, and that's when when I realized, like, hey, maybe the world needs something better, and then we just kind of started the Julia project. So that's sort of we didn't build it for the world. I think that's the most important thing. We built it for ourselves, and it happened to be something uh that resonated with a lot of people, and then it grew into something much larger. And uh, you know, we had no idea, right? Like, like I mean, we we we decided that on the name Julia in like five seconds.

Roopinder 12:51

Uh so tell that story that was just something that was an old project of your professor, correct? And you said, Yeah, that sounds good, something like that.

Viral Shah 12:58

Yeah, it was Jeff actually. So Jeff, Jeff Bizanson had Julia, you know, uh, he had some he had something called Julia that was sort of like a personal project. And you know, we were looking we're gonna start this new thing. We hate MATLAB, uh, we hate Python. Like, we want, you know, the whole two-language story, right? Like solving a two-language problem, right? Like we want something that's as easy as MATLAB but as fast as C. And uh, and uh that's uh or as easy as Python and as fast as C, right? And that's kind of what led to the creation of Julia is that like we wanted something for ourselves that we knew as scientific programmers that that you know would allow us to write large application software. Uh that's how we started Julia. And then uh Jeff had a project, so we were like, Jeff was like, how about we just call this thing Julia? Because I already have this logo ready to go. And we have this beautiful ASCII art logo. I don't know if you've seen it. Um, can I share my screen? Or please, please. Let me let me show you what it looks like. I'm just sharing my terminal here.

Roopinder 14:01

You know, you might want to invent a better story than that because I'll tell you the real one. I understand it can't be your it can't be your wife. Your wife is like we said, has a nice Indian. Well, she's Indian, so meaning that one. Like, could you call it something else, maybe?

Viral Shah 14:19

I I think she's happy I did not do something that tacky. Okay. Okay, can you see my terminal here? Yes, I can. So you just are Julia. Okay, if you just are Julia, okay. You can this uh this logo was checked in you know in 2009 on the very first day of the project.

Roopinder 14:38

Uh you know, this this is the thing.

Viral Shah 14:40

You need I I agree with you, you need a better story, but like I'll tell you what, like, a lot of people on Dreed have better stories.

Chris Rackauckas 14:46

Oh, they do? All right, yes, all right. There have been a lot of invented stories about Julia, that's for sure.

Viral Shah 14:52

Oh, okay. Well, because there's there's Julia sets, right? Like in mathematics. Like, so you have the Julia set, you know, there's a mathematician called Gaston Julia. So, like, there's like you know, so someone thought it's uh you know, the actresses, uh that you know, there's all kinds of there are better stories, there are better stories.

Roopinder 15:08

They are okay, all right. So you're not gonna tell us what the original Julia was, right? That's a deep secret.

Viral Shah 15:15

That's a deep secret. That's unfortunate. We can't go beyond this. I can just tell you how we ended up with okay. The logo I thought about that. All right, very good. Well, it's uh it's uh you know, it's you know, you start these things for yourself and you never know where it's gonna go, right? So you know, it was it's it's years and years. So, I mean, all of us have been doing Julia for like uh almost 15 years now. There's a million users, there are hundred million downloads. Okay, there's a there's a million users, and it's all over the world.

Chris Rackauckas 15:45

There's a lot of people in the United States and in Europe, but also a large concentration in India as well. So it's it's really global. I was also talking about the the the open source contributors, right? Because it's it has a very strong connection to open source, right? One of the big things about Julia is that its package system is made to work just directly with GitHub. And so you have a lot of people that just you know, uh by default, when they first started using the language, GitHub really started getting popular at the same time. And all the code that they're doing was you know not on their own computer, but going straight to GitHub. And so you find that you know there's people all over the world, people who are, you know, in uh in graduate school, all the way up to being, you know, uh late professors, people running, you know, uh engineering departments in different companies, all putting Julia code out there. Yeah, so it's you know, it's it it started when it started, it was like you know, you kind of knew everyone in the room. The the first conferences were you know 20, 30 people in a room or whatnot. And

How Julia Grew Into A Global Tool

Chris Rackauckas 16:40

but now it's uh something where it's yeah, you now you measure it in the you know how many how many million people will will watch this thing if if we put it out there, kind of you're talking about JuliaCon, the con the conference, annual conference.

Viral Shah 16:52

Yeah, yeah. Not just the conference, but like when we make a new release of Julia, you know, we put out a blog post, like everything, you know, it's immediate you and hacker news. There's like people discussing it. It's it's really quite quite a quite quite something. And you know, the the i I remember Chris's point about 30 people. I think when when we uh so so our you know, like I said, we built Julia for ourselves, and for the first three years we didn't actually release it. We accidentally released it actually through a blog post, which you can still find like on the julia language.julialang.org. The oldest blog is the why we created Julia blog. And uh and and Jeff got invited immediately to go to Stanford to give a talk. So he fly flies from Boston to Stanford as a room full of people doing linear algebra and applied mathematics and programming languages, and it was like you know, 30, 40 people that like we're like, wow man, this might be the best thing that ever happened. Like, this is all we needed, right? Like we were like, we're like, if that's all that came out of it, that would that itself was a success, right? Like, uh, but but you know, it was just the start.

Chris Rackauckas 17:55

And then there's a conference in the ASML football stadium, you know, 10 years later.

Viral Shah 17:59

Yeah, right. Like, I mean, ASML rents the football stadium for JuliaCon like in 2024. Like, uh that's a famous one, I think.

Roopinder 18:07

I forget the name, but wow, wow, it really took off.

Viral Shah 18:11

It really took off, and you know, I was joking with uh with Jeff the other day. I don't know, Chris, if you were also around, but um I was like, Julia might be the last handcrafted programming language.

Chris Rackauckas 18:22

Yeah, that that could be. I mean, because yeah, all the others would be AI from this point on.

Viral Shah 18:26

Now on everything's gonna be by AI. I mean, whether high quality or slop, I think everything's gonna be different. So Julia might just be the last handcrafted programming language that the world actually saw.

Roopinder 18:38

Yeah, yeah, yeah. So you're invented for applications, applications, engineering applications, which I'm specifically interested in.

Chris Rackauckas 18:46

Yes.

Roopinder 18:47

And I'll I'll tell you why. My interest in it spurred was spurred by have you heard of this guy, Jensen Huang? Of course. Okay, all right. So he he said he said to us about

Physical AI Beyond Chatbots

Roopinder 19:00

two years, two or three years ago, that this uh agents, AI agents were gonna take over everything. He said, but they're gonna need physical AI. And of course, he gave us his a version of physical AI. So since that day, okay, I thought, wow, you're right. Physical AI is exactly what engineers need because I mean LLMs are nice, you know, they're very nice. They understand language, and we could use them in our apps if they could just be the UI for the apps. That would be great, right? Because they can, you know, they've read all the documentation, the help files, they know how to operate the system. Let's have a natural language interface for for AI program for engineering applications. I hate the fact that I have to memorize the my application's language. I wanted to learn my language, right?

Viral Shah 19:48

Very good point. That's a very good point.

Roopinder 19:50

Right. You know, and and the more like the more programs that you have, the harder it is. Like I use ANSIS, I use SOLIDWORKS, I use several CAD programs. Each one of them is its own 300, 400 words, right? And so I can't be a master in all of them, but I do want to say things to anyone that I use, whether it's ANSIS or whether it's a design program, design me a bracket or analyze this bracket, right? And I want it to just do that, right? I I don't want it to have to learn all the tools. Okay, so okay, where am I going with this?

Viral Shah 20:25

I gotta but that's the LLM. I think what you're saying is that the LLM world, but there's the physical AI world which goes beyond the LLMs.

Roopinder 20:33

Yes, yes. And the fact that robots can't handle the physical world, right? Which was I think Jensen's main point is like he's gonna use physical AI, so his robots are gonna be able to work in the real world. Because with physical AI, they can sense the physics of an object, just the fact that it has mass, that the robot's hand shouldn't go through your hand when it shakes it, right?

Viral Shah 20:57

Or even worse, it should not crush your hand.

Roopinder 21:00

It should not crush your hand, right?

Viral Shah 21:02

Like would you would you shake hands with a robot? Like a lot of people would not give it a thought. I'm I'm like, I'm like that thing can literally crush all the bones. You know, it has it, you know, it has enough strength, right? It has it has enough power to do that. Absolutely, absolutely.

Roopinder 21:16

My first thought. I I gotta say, we're all we had the my old company was called engineering.com, and one of our graphics that we used was a robot hand holding, shaking hands with a human hand, right? And every time I saw that, it was that logo was made before my time ago, I would never shake that robot's hand, right? Because I don't know, uh it can't sense it, has no physical sense, right? It doesn't understand physics or even contact, right? Yeah. Okay, so now how are you gonna use diet, your project? Am I correct in calling it a project? It's not a company, right? Or is it a product?

Viral Shah 21:53

Diet is the product, so Julia Hub is the company. Okay, Julia is an open source project that many.

What Dyad Models And Why

Viral Shah 22:00

Of us who started the project also started the company, and it's a big part of what the company does. Um, so for example, some of our customers like Boeing and ASML use Julia for their control systems, and and that's you know, that's uh that works very well for them. Uh but then you know we kind of keep moving up the level of abstraction, right? And so that's where dyad comes in, where you know, now we have a domain-specific language that sits on top of Julia for solving complex engineering problems in systems modeling. And you know, if you're familiar with systems like Modelica and Simulink, you know, that's you know, you you mentioned a lot of ANSIS and SOLIDWORKS and CAD, right? Those tend to be like component level modeling tools, right? Like if I want to, you know, model like the perfect shape of like the cylinder of my engine or something, right?

Chris Rackauckas 22:49

Right.

Viral Shah 22:50

Um, or or like the perfect shape of the airfoil. But like, you know, if you're if you're like actually designing a plane, right, you need to model all the systems and all the subsystems. Those things are usually done in a systems modeling, multi-physics systems modeling tool. And that's what we build, right? So everything from you know, the you know, figuring out the um the aerodynamics to the avionics to the attendant call button to the lavatory to the uh you know the the air conditioning system, you gotta model each and every single one of the, you know, the the braking system, the landing gear, the flaps, every actuator, right? And typically that's done with a systems modeling tool, and that's what dyad is, right? So it really makes it easy to sort of express the physics that is inside of your uh product and at all levels of the product and uh and simulate it before you even need to build it.

Roopinder 23:45

It's coming back to me now, Chris. Your your demo where you're showing actually how to make a car. And I gotta tell you a little story. I was at a conference and they it was okay, it was on a desk. It was, and they they had a they this was also about two years ago at Dust University, their big user conference, and they introduced a program that was essentially AI make me a car, right? And as an engineer, I thought this is absurd, right? Because as I rolled out, sure it was getting the outer shape of the car, right? It was nice looking graphics over it, showing the wind flow or the car, but that was all superficial. There's nothing inside the car, there was no system, right? There was no braking hydraulics.

Viral Shah 24:33

I have a demo for you, Rupidra. I have a demo for you if you don't mind. I can pull in a colleague that'll say, Computer, build me a plane, and it'll actually build out the control systems for the plane.

Roopinder 24:44

I would love to see that.

Viral Shah 24:45

Would you have to just run a presentation for you? Like I'm just gonna see if I can find him on Slack. Uh, just give me a second.

Roopinder 24:51

All right, okay, that'd be great.

Chris Rackauckas 24:53

But yeah, I mean, the you know the the the big thing, the big thing about it, right, is you know, if you want to get to that physical level, I mean what you want to start with is you want to start with a program that does high fidelity modeling and simulation, right? Because you want to be able to have the whole high fidelity physics of the world and add machine learning to it. I think a lot of folks kind of take a look at this problem, they just say, hey, look, I got a machine learning tool over there. Let me just try to put a little bit of physics on it. And physics is a lot, right? There's a lot of detail that goes into making it very accurate. So instead, you have to really focus on getting a very accurate physical engine that can handle all the details of everything that happens about contact, everything that handles about you know these high-dimensional kinds of problems, and then say, now machine learning, do the simple thing, you know, do do do do as little machine learning as possible, but fix things to make things more realistic, right? Because when you when you would you do physics, you you make assumptions, right?

Viral Shah 25:47

Chris, the the known physics and the unknown physics, right? And you need to model both, right?

Chris Rackauckas 25:52

Exactly, right? You start with spherical cows, then you make them you know less spherical, and then you you know you you say, Oh, all gases are ideal gases, right? And you, you know, and so the engineers really work to make something as realistic as possible. And then you say, Here's my data. Now, from the data, learn how to make this extremely realistic, right? Right.

Roopinder 26:10

So learn from the data and learn from engineering principles. Take what I know already, take what's established known as right. My I'll say wisdom for lack of a better word, but learn learn go from where I know, go from the systems that I trust, right? I trust some of these products, software packages that I use, but don't go throw them all out and say you're gonna make me a car based on I don't know what. If you make me a plane like that, I won't fly in it, right? It's not gonna be real, right? So I love the fact that you put the car together when Chris said your demo, you're actually putting a car together by its individual systems, right? Yep from known things that worked, systems that worked, right? Which is which has always been my my frustration with AI companies, like startups that that start from scratch and say, okay, I'll make you a building, I'll make you a room, or you know, and they they're often very focused on the physical form of how it looks, but not the physical, like not the function, right?

Viral Shah 27:14

And and you know, liked is all focused on function and not form. So we don't even deal with geometry, we work with the systems, and we can leverage all the other tools around geometry, right? Like there's so many CAD and CAE tools that are out there, it really doesn't make sense to build another one, like we should just reuse, and there's so many open source ones, so many nice commercial ones. We have a partnership with ANSIS, for example, with Synopsis. Yeah, so so we can pull it all together. I have my I have my colleague David, uh, who can uh show you a quick demo of you know of the HL20 spacecraft uh building it from spec. Okay. So so hey David.

Speaker 1 27:52

Hey David. Hey, how are you? Nice to meet you. My name's David.

Roopinder 27:55

Nice to meet you. Thanks for jumping on. You're coming. Yeah, yeah. What's that what's that feel like? What's it feel like when the boss says, Hey, I need a demo, like right now.

Speaker 1 28:07

It feels like I'm useful and then I have a job.

Viral Shah 28:10

David can David can whip out demos like with five seconds notice.

Roopinder 28:15

Oh, okay.

Viral Shah 28:16

All right.

Roopinder 28:16

So David, no pressure, really. So it should be edited.

Viral Shah 28:20

So we're gonna, if you uh if you get a as David is setting this up, right? So this is this is uh you know the HL20 lifting body, and it David will say what it is. But like we're really thinking about like we've got this agentic stuff, right? And and you know, you got cloud code, and you can tell it what software to build, and it will help you build it and write the tests and build the GUIs and everything. The question is, why can't you do that for hardware, right? And then in hardware, you have sort of all these click and drag and drop tools, and you're like, okay, but like why can't I have cloud code for hardware? And the answer is you can have cloud code for hardware if you have dyad, right? Because dyad is a domain-specific language, it has the full access to the Julia environment, it has a dyad compiler, a full understanding of the physics. And so it can put all of these things together in the form of a program that you can then simulate and you know, basically design real hardware while being inside an agent. And you can now start asking the questions that oh, people are doing, you know, people are using open cloud to run all these agents to do all these things on the software side or have personal assistance. Well, why can't I have personal engineer design a plane for me? Right. And so

Demo: Building The NASA HL-20 Model

Viral Shah 29:30

that's uh let's have David run this, and this will drive a lot of conversation, I think.

Roopinder 29:34

Okay, great.

Viral Shah 29:35

Are you able to see my screen with me drawing it right here?

Roopinder 29:38

I can see it, red squiggly red line.

Speaker 1 29:40

Okay, great. Yeah, so I'm gonna talk about um how we use our uh diet agent here to design diet model of the HL20 lifting body. So just a little bit of brief background on what this thing is. So the NASA HL20 um is a 1990s, early 90s NASA concept vehicle uh designed to transport astronauts to and from um low Earth orbit, specifically um the space station Freedom, which is one of the precursors to what eventually is now the uh International Space Station. But the idea is this was a lifting body, no large wings, the body generates the lift, but it was never built due to shipping priorities, change in budgets, and so on. But the reason why this is a very interesting model for us is NASA published extensive uh documentation and aerodynamic data on how to simulate this model, right, with wind tones testing data, fight, data from programs, and CFD. So, what we are trying to do with this is we are doing a proof of concept to demonstrate that the diet agent has the ability to build a model from scratch, right? Given a specification, can it build it to specification? So, here what we're doing for this diet agent is we're giving it three inputs. Uh, these two right here, these are technical memorandums from NASA, which you can see right over here on the right hand side. Um, and this is the uh the data that um NASA published with um with this model. And also we're gonna give it a DML of lookup tables for the aerodynamic data. Right? So traditionally, when uh if you're working with a from scratch, you might be handing a junior engineer or group of engineers these documents, right? These are PDFs which are uh human readable but not necessarily machine executable. So there's diagrams, text, paragraphs of words describing what it is. And then we also are providing a DML file. So if you usually give this to a junior engineer, it would take them anywhere from weeks to possibly even months to build up this model by scratch, right? Reading through all the documentation, making sure that the models, the coefficients, and the equations are expressed the correct way. So in this uh proof of concept, what we did was we fed all these things instead of to a junior engineer, we gave it to the diet agent with the corresponding prompt. And what we got after about 40 minutes of the LLM moving along is we got a diet model of the HL20 fully built by the agent. And what we're doing with this model that we're building that we got from the agent is we're we've edited it by controls and aerospace engineers to establish a source of truth. And then we're using this golden model, this source of truth.

Roopinder 32:12

Yes, but you say uh you're gonna go into this, right? But okay, fully built. You're gonna go into fully built, right?

Speaker 1 32:19

What that means, like how I'll I'll show you the actual model itself. Okay, all right, okay. Yeah. So yeah, so I'll just go ahead and skip to that right now. So this is just a little preview of the so we it took, I said it took about 30, 40 minutes for the agent to do it. So this is just a really quick screen capture preview of the of the agent actually working along. So on the left side is the diet agent, which would be given the prompt and the three inputs that I just mentioned in the previous slide. And then on the right side is just vanilla claude code. And claude this in in this instance, Claude Code was also giving the prompt the same inputs, and we also gave it um access to diet tooling, which is Julia REPL and the diet compiler and so on. But the point here is the diet agent was able to completely build the model that compiled, whereas the claude um was not able to do this. It actually compacted or ran on memory about three or four times and eventually ended up crashing. Yeah. So this is it, right? Same task, diet agent was able to do it, Claud Code was not able to do it. And this is actually so this is one of the models that the agent actually built. So if we go into the PDFs, one of the more the most complicated um components that the agent actually needs to create is this HL20 mixer model. And this mixer model is a control surface mixer. What it does is it converts pilot commands such as elevator, pitch, speed, and control um into actual deflections on the surface of the HL20 itself. All right. So I'll go ahead and give you a little preview of pull up the um.

Roopinder 33:55

So just to clarify, when you say making the model, you're talking about a systems model that it's a system like their system diagram that you're showing us. You're not, are you you don't get to the point where we actually make the physical model or a 3D model of it, correct?

Speaker 1 34:10

You know, or is that we don't no no 3D model, right? Right, Chris?

Chris Rackauckas 34:14

Yeah, yeah. Yeah, so this is a model of of the of the physics, right? I mean, as a as a plant model, but that's uh captured as a system model, right? So that's like at a higher level, uh it's uh, you know, it's at a lowered level fidelity than having, you know, like the whole you know, wind going over. Now there's a lot of uh actual information that's captured, right? The DML files that were talked about, that's aerodynamic data to be able to know like, you know, what are the aerodynamic drag forces of the body when it's at different angles, right? So it has a lot of this data captured from the actual device as lookup tables, and then it uses that that look those lookup tables to know, okay, if I'm at this angle, this is the drag, this angle, and then then it is to be able to build out a plant model that is a model of how, you know, the higher level, how does the whole body move in general? And then we also model things like the control system and all that, uh, so that way we can get a model of the whole system, right? And that's kind of the key about system modeling. If you if you if you do a lot of this, you know, oh, if you want to do the extremely high fidelity, you know, look at the airfoil, look at the the, you know, look at the the drags going over it, then you can basically only look at one wing, right? You know, you see all those CFD things. They look at one wing and they do this simulation for like, you know, three days, right? And so we have to say, okay, we we we get some data out of that, and then we collapse that down into data tables, and then we build out the whole system. And we need to be able to simulate the whole system very rapidly because now we're going to want to put this into real-time context. We're gonna want to, you know, have a controller that uses this information in a real time as it with real-time gains. So you do have to lower the fidelity of the model a bit, but then there are ways to be able to keep some some uh real reality to this.

Roopinder 35:46

Okay, so there's this different way, a different way of looking at the aircraft than I'm accustomed to, but I totally get it on the system level. You could use it, you could use it uh for input and output, like if you vary the uh mission parameters, for example, like lift different uh orbits, or I'm thinking I'm not space aeronautical engineer, but different orbits, heights, launches, angles, uh, all that stuff. You could you could we could vary the input and see what happens to it? Like a similar thing.

Chris Rackauckas 36:17

Exactly. Okay exactly. So you can you can do things like uh understand, you know, if there's uncertainty in the wind, um, how likely is is a is a landing to be safe, right? Um that'd be very difficult if you just were looking at one airfoil and scaling that all the way. But here you can say, okay, I can vary all this. And and you know that there's a certain level of error that's introduced by it, right? But you know, if you capture enough of the of the of the physics, you can start to get some pretty good uh ideas for the uncertainty. And then this leads to, you know, design uh higher level design decisions, like, okay, you know, is this a safe design? Um, is this a design that we could be using if we're using this in Florida, right? Like all these different, all these different higher level concerns can be answered by going to a slightly lower fidelity model.

Roopinder 36:58

Got it. It's all built on the understanding the documents that you documents and uh specifications that you provided it, right? It can it read those and it can understand those and it's just processed them all to build this model correct, this systems model.

Chris Rackauckas 37:14

And maybe show a little bit of the diad code where you can you can see like you know, it's modeling things like the rudder and okay, okay, all right, all right.

Viral Shah 37:22

Yes, if you there's a one-to-one correspondence between the diagrammatic view and the code view, and so you can go from one to the other without losing any information, and you can use an agent to then modify both of these simultaneously. So that's sort of the new paradigm of how the system works. But you know, here you can kind of get a feel for like what the code looks like, right? And you can if you go a little bit below, then you'll also see like some of the equations.

Roopinder 37:48

Okay. So you can actually see those differential equations, huh? Okay, spare me, spare me those, but I the code is bad. The code is bad enough, I gotta say, but I don't want to go any level deeper. But so I don't have to write the code, right? I'm playing, I'm just dragging and dragging the blocks. Okay, well, this was all generated.

Viral Shah 38:08

This was all generated from the aid by the agent by reading the P.

Roopinder 38:12

The agent generated that block, the blocks that I'm looking at. Yes. Okay, yes. But if I but if I want to change something, I could change, I could the interface is through the blocks, correct?

Speaker 1 38:22

I could yeah, you you can you can change everything on the GUI. You can you can manipulate it through the GUI on the left or right side, and the changes will flect and both sides. Yeah.

Chris Rackauckas 38:29

Well, and so so it's it there's a one-to-one correspondence. So if you make a change in the graphical user interface, then it makes a change in the code. If you make a change in the code, it makes a change in the graphical user interface. And it goes bi-directional and it always stays in sync. Um and that's kind of key because you know the agents and some programmers like to work in code, right? But then you have a lot of engineers that like to work in the graphical user interface. And so by moving to this kind of two-way form, but by developing a language that also has a direct representation to visuals, um, we can touch both communities directly without having to have a compromise between them. Got it, got it.

Roopinder 39:05

So I could do systems level design of whole systems, even as complicated as a spacecraft here, the heavy lift spacecraft. I'm looking for my I'm looking back here for my systems engineering textbook, right? I'm not I've I found no use for it, so I may have thrown it out, but I may have to dig that up. But I got my stress analysis books back here and all of my some CAD books. Uh, but but yeah, okay. I'll have to brush.

Viral Shah 39:36

Actually, on that topic, just uh very quick, David. I just uh sent you a link. If you could open it up.

Roopinder 39:42

Oh, and David, we interrupted your demo. Is there more?

Viral Shah 39:46

Well, there's a lot more actually, but we'll we'll get to it quickly. I wanted to

Validation Plots And Trustworthy AI

Viral Shah 39:50

point this thing out to you. Um, you know, so you're talking about CAD and geometry, right? This is your typical sort of engineering design process, right? I'm sure you've seen this V before.

Roopinder 40:00

Yeah.

Viral Shah 40:01

But you start from the system requirements and then you go to you know system level design, subsystem level. And then when you really get to component level design, that's when you start doing all the geometry work with CAD and like what's the exact shape of this part and what is the stress, and you know, is it is it going to be able to sort of survive this heat, or like if you're doing electromagnetics, if it's a communication system. So you do all the detailed work at the bottom here, including the software, and then you integrate it into subsystems and systems, and then finally you might deploy it and you might even have digital twins for predictive maintenance and runtime optimization. So DIA kind of sits on the upper arms of the V here, whereas you know, all the tools that you see from ANSIS, Cadence, DASO, Siemens, all of those are you know the CAD tools, the PLM tools, all the 3D simulation solvers, they all sort of sit in this component level analysis. And uh, you know, they tend to sort of model a single PDE, you know, at a time, right? Like each, like Fluent is all about CFD, right, for example, or ANSIS mechanical is all about structural. Um, whereas the tools that you see up here in the design and the operational phases, they're multi-physics tools. So you can combine, you know, stress equations with electrodynamics, with uh, you know, with thermodynamics, with uh, you know, uh, with fluent mechanics, whatever you need, right? Whatever physics you need to bring in can be modeled at in the systems modeling capabilities. And I think the point you made about core being hard is literally what is blocking people from doing some of these things, and the agents are gonna change that. So you're you know, every engineer learns systems design and systems modeling at the end of the day, right? The question is how many of them actually use it in their day-to-day jobs? Like, unless you really have to work on something sophisticated like automobiles or aircrafts, you kind of tend to just kind of wing your way through it in the lab. And if you make the tools easy, you'll be able to design more on the computer just like you can with CAD today.

Roopinder 42:01

Yeah, yeah. Oh, it's perfect good use of AI because it takes something I I don't want to do, systems modeling, and put puts it out there where I could actually uh have access to it.

Viral Shah 42:13

And if you want to ask questions, or you want to design and ask questions, say Chris, go ahead.

Chris Rackauckas 42:17

Yeah, and then effectively, you know, and a lot of people who might not be very good programmers can interact with it and ask questions and say, okay, you know, you know, like some of the high-level questions that that I said, like, you know, what are the safe win conditions for for this uh landing body, right? You can type that into the agent and it can say, hey, look, I created these different scenarios and here's the plots for them. And you can subtract the plots and you can say, oh, you know, I you know, you you still know enough about engineering, you can look at plots and say, oh, that's safe, that's unsafe, that's okay. But you didn't have to write all the code for checking all the different scenarios and for trying everything out, right? You can really just interact in human language, you know, to be able to get something that is then under the hood running actual physical, you know, and systems level simulations and then giving you back plots for that instead of having to go to it to that deeper level.

Roopinder 43:02

Okay. Okay, I get it from the systems engineering point of view. Now, when it gets to the the next part of the V. So now when it gets when the systems is system is okay and I'm happy with it, now I can call in the design engineer who can actually model it. What kind of data is coming across? What kind of data is being output from Dad to help this design engineer? What does he actually have at his disposal? Because he's not going to get any kind of sketch or anything, right?

Viral Shah 43:31

So you actually maybe this is a good point to continue with the demo and you'll see what comes out. David, if we can look at those plots, right? I think that that might answer some of those questions.

Speaker 1 43:43

Okay. Yeah. So one of the things that's um inside of the PDFs, the technical memorandums from NASA, is they have dynamic check case plots inside the documentation. So what that means is there's some specific trim conditions that that NASA's documented. So in this case, We have a trim case zero, right? So at a certain condition, the HL20 is supposed to behave like XYZ. So if we kind of go down here, we can see that there's a test case here for this trim case zero. So we have a pitch pulse test case. What that means is the eight when the HL20 is at trim case zero, if the pilot inputs a little pitch maneuver into the HL20, how do we expect it to behave? And then now there's actually these validation plots provided for us inside the NASA documentation. So right, so right here we have a single pitch maneuver from the pilot, and then we're able to see how the HL20 is supposed to behave from there. So one of the things the agent actually created for us is this right here a test from Appendix F, which outlines the pitch pulse test, which is the test I'm showing you inside this documentation right here. So what I'm gonna what I'm asking the agent to do, and I I went ahead and asked the query while you were having this discussion, but I asked the agent to please run the pitch pulse simulation for me and then create a three by two summary plot. The things I want you to include inside the summary plot are the uh pilot's pitch pulse um maneuver, the pitch rate, angle of attack, elevator command, speed break, and so on. So here we're asking the agent to actually simulate the task that it created for us and then plot the variables um that we're interested in. So I'll kind of really it's gonna do that. Yeah, so it's gonna work through this right now. Um I'll it actually see it goes into here, it it reads through the documentation, it reads through my repository right here. Right here, it actually goes into the pitch pulse died file that that outlines the um the test itself, and it's reading all this and it's reading to understand. And if we go all the way to the bottom right here, we can actually see it running Julia code. So if I expand this right here. This is currently live, right? It's it's been running for three minutes and 30 seconds since I asked this, since I asked this uh query. So it's it's running this these Julia commands right here. You can see that it's bringing the the the project into scope, and then it's just running that test app f pitch pulse simulation that we asked it to.

Roopinder 46:15

Right. So it's dyadic. I see it's run up uh four and a half thousand tokens. Is that a lot? Not really, not so much. Yeah, I have no idea.

Viral Shah 46:26

Let me say this. If you had to do this manually, you would have been going at this for about an hour, like a proficient person, like a highly paid engineer, just kind of you know, to do this one test would have spent an hour.

Roopinder 46:38

Okay, okay. Yeah, no, I I believe you.

Viral Shah 46:42

And you know, it it created something which did not work and then it fixed itself, and and then there you go.

Roopinder 46:48

So just this is a it just created this right here. That is amazing. That is amazing. Can I trust it? Let's find out.

Chris Rackauckas 46:56

You can also go back and also look at all the code, right? And that that's why you know it's it doesn't, it's not just a black box, it's a it's a system that's actually you know creating all the artifacts to be able to say, hey, to recreate all the you know, to be able to recreate anything. Here's all the code.

Roopinder 47:10

It's built on the governing equations that would rule here. That's uh the physics based on the physics.

Viral Shah 47:15

It is. So David, can you show show can you show the PDF which has the validation test cases? Yeah, right, because I think that that's in this case, this is the validation, right? In the real world, the validation would be with experimental data or other sources.

Roopinder 47:31

Yeah, this slide reason I'm asking is because I have been I've been burnt by uh burnt by AI many times, right? It presents me a solution I think is like amazing at first, right? Wow, this would have taken me literally days to put this to the report together. And then upon checking it, I find uh oops, that's not what I would have said, or that's not quite right. So I find myself in the process of spending, I think I'm probably not spending as much time checking it as I would in creating it, but it feels that way. Do you know what I mean?

Viral Shah 48:08

No, you're you're absolutely right. And this is where sort of the agent is, you know, the agent is a generative part here, right? So based on reading the docs and the PDFs and everything, it creates a model. But at the end of the day, that model has to validate, you know, has to have physics that are consistent inside, right? It has to be on the you know, based on the governing laws. And if there is an error there, then the compiler, the diad compiler will tell you, hey, you had a you know mass uh conservation uh violation here, you need to go and fix your equations, and then the generative part of the tool is gonna go and you know fix that until the model works, and then once it works, you want to validate the test cases, right? And this is what David was just showing here are the test cases, right? So on the left is what our agent created, but on the right is what you had in the original PDF as the validation test cases.

Roopinder 49:03

Uh, I see. Oh, I see, I see. Okay, okay.

Viral Shah 49:05

Right, you can kind of see that they've been scanned like tilted, right? Like it's not even like a real like PDF, right? Like it's a scanned PDF.

Roopinder 49:12

That looks damn good. Uh first inspection.

Speaker 1 49:16

Yeah, so what you're looking for here is that the um the shapes of these uh look similar. There is gonna be some slight differences in the plots, and the reason for this is inside the NASA PDFs for the various trim cases, they didn't always specify all the initial conditions. So some of that was um to the question, the engineer um setting up the these problems, what the initial conditions would be. Um so basically lack of information about initial conditions is what's leading to these slight differences. But but yeah, what you're looking for is mainly the shape and the behaviors the same.

Roopinder 49:47

Yeah. I got it. What do they call that when you uh drag in a professional on your team and they say like a professional basketball player or something? You're not David, I'm asking because are you like an aeronautical engineer they just happened to bring on here, or did you have to learn aeronautical engineering for this?

Speaker 1 50:05

My background is in aeronautical engineering. Um so it's my undergrad degree, but that is uh only one part of small part of my responsibility here. Yeah, I would not have thought of that as the reason they brought me in, although I would be flattered.

Viral Shah 50:20

It turned out to be useful.

Roopinder 50:21

Yeah. I would say so. What's the word I'm looking for? What's what do they call that when you bring in a pro? I'll think of it for this show, of course. The validation is is very, how should I say, necessary? Uh, because you can't just ask the AI to validate it, right? You have to validate it against tests or basic physics. You know, we do like back of the envelope checks to make sure that's within our solution that ANSIS gives us, for example, is in the right ballpark.

Viral Shah 50:55

This is where the compiler that enforces the physics is essential, right? And now you can start asking the questions, like David said, right? This was an underspecified model. So now you can start asking, okay, like what can I learn from the data that is not present in my models that can allow me to make uh you know, yeah, the uh the next, the second order of improvement. And you can kind of start leveraging data to improve your models as you go along. As we talk to our customers, right? Like one of our largest customers is a major aircraft design uh manufacturer. And you know, the engineers they have they have thousands of engineers who are working with data like this, right? And with designs that that look like the demo I showed you, right? They're not sitting in CAD or like they're not doing physical AI in the way sort of like you know, people in the popular consciousness, right? Like, or the larger public thinks about, you know, when they think about AI or when they think about manufacturing, right? It's all this sort of detailed subsystems design, right? Like, what does my electrical look like? What does my aerodynamic system or my flight controls? What you know, what are my safety systems looking like, right? Like uh, you know, what does the environmental control system look like? Like each of these things at every level of detail, someone's sitting through and working test case by test case by test case and making sure it's correct, and that when they bring these things together, the larger system is working correctly, and then you you know build it all the way into a plane and you want to load it into a simulator that's not a flight simulator like like you see in you know, like in a video game or an arcade,

The V-Model And Electrifying A Car

Viral Shah 52:30

right? But is a flight simulator that will actually fly the plane in software and verify that every single physical quantity is kind of what you need it to look like. Very few people are actually working with the physical form of it, right? But thousands of engineers are usually working with uh with the underlying systems.

Roopinder 52:49

I like the idea that it's using real physics and the governing equations. Uh the interface, though, it that's the only thing that gives me pause. It's like I still have to learn it though, right? I still have to learn it.

Viral Shah 53:04

Oh, you all you needed is the is the agentic interface.

Roopinder 53:08

Oh, okay. Okay.

Viral Shah 53:09

What what you're seeing is what the agent generated. You did not have to do that. Oh, okay. You can if you want to, right? If you are an old school engineer, like I don't trust this AI computer stuff, I'm gonna not do it unless I did it by hand and you know wrote down the equations on my whiteboard and verified. Yeah, you can do that because you have to be able to do that. And the agent can only be successful if a human could have been successful, right? I mean, at some level, the agent is like a fast typist, right? Like it typed, it all it typed out the perfect program for you. And so it's important that you have a system that works from the first principles where you could by hand have done all of these things, and then we train the agent to learn all these things, and now the agent goes and does it. But you, as a user, all you did was you took a specification, you uploaded it, the system chugged away for an hour. Maybe if it was more complex, it might have taken longer. And uh, you know, once it comes back, you can then inspect it, inspect the test cases, you can add more test cases. In the real world, specifications of new products may not be as complete as an old NASA spacecraft, which has been studied well. You know, there'll be a lot more to it, but this greatly increases the velocity with which you can have engineers work on it. And instead of sitting and clicking and typing, you know, you're actually doing real work. Um, you know, instead of sort of in some sense like filler work, right? Like if you see what I mean, like I'm now talking about the functionality and the productivity and what does my product do instead of going into all the little nitty-gritties and uh you know, writing, writing little bits and pieces of everything.

Speaker 1 54:48

Yeah, a good example of that is how we just got the agent to create the plots, right? Right, writing the just the Julia code to format the plots correctly to make it look nice, have the variables in the right places. That that would take someone like if someone's really, really good, maybe like 30, 50 minutes, someone, someone else, like maybe an hour or two, but you know, you you don't have to know how to use Julia. You don't know how to need to know how to use diet. You just can ask the agent to do it, and it did in about four minutes for us. So that's an example.

Roopinder 55:14

Tell me if I'm oversimplifying the because this is the way I'm gonna try to present it to our audience of design engineers. Basically, I'm taking all my specifications for the contract, right? All of the all the physics that I need to know, whether it's textbooks or handbooks or something, I'm feeding this into dyad, and the result I get after it processes all my document documents is that system diagram. Is that is that accurate?

Chris Rackauckas 55:43

Yep.

Roopinder 55:44

I give it a few natural language prompts that say something like, okay, now I'm way oversimplifying this, but make me that airplane, right? Or make me that airplane based.

Viral Shah 55:57

You know, here's a specification, build me this airplane, right? Like literally like that.

Roopinder 56:02

It can't be like that simple, right?

Viral Shah 56:04

In in in some, you know, you know, if you had to put it in one line, that's what you would say.

Chris Rackauckas 56:09

Okay, I mean it's yeah, but uh that that's the design phase, right?

Roopinder 56:12

I think it's yeah, well, I'm thinking more in in in degrees, more in uh stages, right? If like if I'm electrifying a car that I already have in my product line, right? I I want to just the only difference is it has an internal combustion engine now, but I want to make it electric, right? I want to electrify it. I want to say in stages, let's look at the motor. Let's change the IC to a to an electric motor, right? And put in performance characteristics of that motor into diad. It can now tell me how it's gonna affect the weight distribution and it's gonna affect this handling.

Chris Rackauckas 56:46

Might be a good time to show the V again.

Viral Shah 56:48

Okay, yeah, so David, can you pull up those? Yes, I will find the V.

Speaker 1 56:53

I got it somewhere right here.

Roopinder 56:55

Electrification of a product, it seems like that's been done a hundred times. I think AI should be able to help me with that. But I it should help me with that. All I did is different subsystems inside a vehicle.

Viral Shah 57:06

Well, and then and you know, as you as you sort of do the different subsystems, right? Like, you know, you you might have replaced the ice with a battery, but now your battery is being used to sort of you know uh heat the engine when it's the dead of the winter in Boston, right? And now your range is going to decrease. So maybe now you need a bigger battery, but now your charging time increases. So you need a you know high kilowatt charger. Exactly.

Roopinder 57:27

I would play with all those parameters. I would play with all those parameters before I give it start designing it. I want to see how all the different uh yeah, all the different ways that one's one thing subsystem plays off the other interactions. Go ahead. Yeah.

Chris Rackauckas 57:40

So so then with that, you do something like so you say, okay, here's my internal combustion engine, and you know, here's everything I know about it. Um, and also feed it all the data about all these electrical components and now uh take this system model and make a version that that seems like a good specification for an electrical model, right? And I'll take that and I'll create this whole systems model and I'll find out, oh, you know, one of the things that's gonna be an issue is that when you turn on, you know, you need to make sure that you have a high enough, you know, uh wattage so that way, you know, when the when the air conditioner turns on, it's not going to short the whole. You're not gonna so we need to have some some buffer here and here. And so it's gonna reconfigure the whole system and try to come up with that system design, right? And that's the the design part of the of the V, right? And then you have to go to the component designers and say, can we build an HRAC system that matches these specifications? Can we build a drivetrain that matches with this with this amount of model? Can can we actually do this, right? And they all design all their components and they'll get close, you know, it won't be exact, right? And then the next thing that comes out is once you have all those components, you know, all those you know, CFD models, all those answers thing, right? You get all that data and you just say, okay, well, here's what we actually get for these components. Here's what we can actually physically build. If we take these real parameters and put them back into the model, does the thing still run, right? Uh and so that's in that that's in that that uh deploy phase, and that's a digital twin where you then say, okay, you you take the model that you built out and you do some machine learning or some analysis on the data out of these ANSIS models to feed it now back into the model. We say, well, this is what we're actually getting. Does this car actually run? And the hope is yes, but if it's no, then you have to go back to those components and say, we got to figure out something different here, or we either need to change, you know, how whether whether we, you know, how quickly we can cool down the car. So we have to change some of our requirements, or we need a bigger battery. And yet, you know, but you can do all this iteration directly in a computer where most of this is happening, you know, typing out in natural language, just trying to find out how how should these specifications be be improved, right?

Roopinder 59:38

Sure, or how I'm gonna change the suspension system because the weight distribution is now way different, right?

Viral Shah 59:44

And uh yeah, and by the way, what happens when you're going uphill, right? Like you're on a slope and everything's different now, right?

Chris Rackauckas 59:51

Right.

Viral Shah 59:52

And by the way, so if you were move in your example, moving from ice to uh a battery, now your braking system is also completely different because now you're gonna move to regen, right? And and so now the question is how much energy am I gonna be able to recover, right? And and so it's you know, it's a series of different different things that you can do. You could you could do this in a very sort of classroom style at a very high level, but you could also you know use the same tools that we build and actually do a commercial level system, um, but you just have to spend a lot more time working on each subsystem then being very clear about the hope is that I would learn from other electrifications that my company has already done and use those as guidelines for the next model to be electrified, right?

Roopinder 1:00:38

That's what I find lacking in

Bringing Your Company Context Securely

Roopinder 1:00:40

AI systems at the moment is like they don't take into cont they don't take the context of my environment in into my environment being my design environment. Right.

Viral Shah 1:00:52

So the big problem that that has been, you know, and that that the world is solving now, right, is that what you really want is an AI system that has access to your company's internal designs and documents, right? And I have a guarantee that those will not leak out into the cloud and go and train some other LLM. And today, you know, if you you know, Anthropic, AWS, Microsoft, all of these guys, right, will give you those guarantees that this, you know, that their models are only going, their LLM models are only going to look at your company's data and never gonna upload anything back into the you know, outside of your company's network. And and so when you're working with DIAD, just like we gave it a NASA PDF, right? We could have given it 20 other PDF documents and said, okay, this is my context, right? Like and now, based on this, let's start working on the 2025 Rev4 plugin hybrid, right? Uh, you know, you have the 2024 one, let's build a 2026 one now. And and oh, by the way, this is what I heard from the markets. I want to make these, these, these changes. Oh, and these new technologies became available, so we're gonna tweak the design in this way, and then you you know have updated requirements, you run your simulations over and over again, and then you finalize it and you know, send it to manufacturing or customers, like in our in our mission, right? Like, we don't want to build a point tool for one industry, we want to build something that is gonna broadly change

Closing Thoughts And Listener Outreach

Viral Shah 1:02:15

the way people build things across industries. Sure.

Roopinder 1:02:19

Yeah, makes sense. I'm impressed, gentlemen. Thank you so much. Thanks for your time. It's been great. Thanks for being on the show. I hope to keep up with uh what you guys are doing. Definitely great conversation and great demo. Thanks, thank you all. Thank you, thank you, Rupinder, and thank you for listening to Faux Des the Future of Design and Engineering Software Show, brought to you by Enge 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 Rupinder at EngeTechnica.com or message me on LinkedIn.