AI for Supply Chain and Design — with C-Infinity and Hestus

Episode Summary

The episode delves into how artificial intelligence (AI) can revolutionize supply chain management and product design through the work of C-Infinity and Hestus. Sai Nelaturi from C-Infinity specializes in CAD and digital manufacturing, aiming to bridge the gap between design intent and manufacturing execution using AI. His company focuses on converting CAD assemblies into production plans with an AI-driven approach. Meanwhile, Sohrab Haghigaht of Hestus brings aerospace engineering expertise to automate repetitive tasks in hardware design and manufacturing, particularly focusing on constraints and geometries in 2D and 3D designs.

Key insights from the discussion highlight the transformative potential of AI in streamlining manual processes typically handled by engineers. Both companies emphasize that their technologies can significantly reduce errors while allowing designers to focus more on creative aspects of their work. Sai and Sohrab also discuss the broader implications of AI, suggesting it will not only change how we do things but also create new job opportunities as traditional roles evolve.

For PLM and engineering professionals, the episode underscores the importance of embracing AI technologies to enhance productivity and innovation in design and manufacturing processes. By adopting these tools, professionals can better manage the complexities of system design and assembly, ultimately leading to more efficient product development cycles and improved overall competitiveness.

---

Full Transcript

Michael Finocchiaro

Well, welcome to the AI Across the Product Lifecycle podcast once again. I'm joined by two friends, Sai Nelaturi of C -Infinity and Sohrab, and I can't pronounce his last name, Haghigaht of Hestus Two very exciting startups from the Bay Area. Sai, you want to tell us a little bit about yourself and about C -Infinity?

Sohrab

I did.

Sai Nelaturi

Sure, thanks Michael for having me. yeah, so I'm Sayan El-Atoori, I run C -Infinity. My background is in CAD and digital manufacturing. I got my PhD in the space from University of Wisconsin in 2011. I've been working at the intersection of design and manufacturing ever since. And when we started C -Infinity, we were thinking that one of the biggest impacts of artificial intelligence is going to be in how... products are not only designed, how they're manufactured. That interface between design and manufacturing is where a lot of the problems are. And we see a lot of work in design for simulation and in design for manufacturability for the parts, for individual parts. But we wanted to think about the higher level problem of system design and assemblies where parts come together. And lot of time is spent there manually. And at C -Infinity, we're building artificial intelligence that can convert your CAD assemblies into production plans. to go and transform the design intent into manufacturing execution very quickly using artificial intelligence. And so I like to think of what we're building as a compiler for manufacturing.

Michael Finocchiaro

Nice. I love it. That's super interesting because it's very similar to what Sohrab is doing at Hestes, right?

Sohrab

⁓ Yeah, I mean, thank you for having me. Excited to be here. My background is in aerospace engineering, so I've dabbled with ⁓ many things that do fly or are part of flying vehicles from landing gears to my previous company was a space launch company. And I agree with many of the things that Sai said that there's a lot of difficulty in the world of hardware design and hardware manufacturing that I think we should leverage AI. The way we look at this. is there are many manual and many repetitive steps that engineers have to do to be able to bring that vision that they could clearly see in their head into digital world, at least at first, and then later on to manufacturing. And those often involves ⁓ relationship between geometries or as we know them, constraints, and also size of those geometries, dimensions. And we can do that in 2D, we can do that in 3D, and then ⁓ Fly wants to basically take what you have built as an assembly and turn it into a production plan or construction plan. ⁓ We want to do before then that, okay, these are the parts that you have, they need to fit together. For them to fit together, we have specific language, GD &T, but that should carry the intent of the designer. Is it going to be press fit? Is it going to be a loose one? Why is that? And... That is a very manual and error-prone work that designers are doing, and we want to be able to automate that for them so they work and focus on the creative part of the design, less so the manual and repetitive aspects so they can just breeze through it much, more easily.

Michael Finocchiaro

Really cool. I love that you guys are attacking a very similar problem, with two different approaches, actually not even competitive approaches, they're actually complementary approaches, right? ⁓ So on this podcast, I'd like to talk about, well, Jagita, kind put my finger on the pulse of what's actually separating the wheat from the chaff in terms of AI, right? Because there's just so much hype around it. So before you started on the most recent ⁓ work you've been doing, how did you perceive AI? ⁓ You both have a deeper experience. So you probably were using it before, but before 2023, I mean, before chat GPT, we, anyway, so that's sort of the question. It's like what, before you actually started working on this particular product, what was your perception? Did you think this is the new frontier or maybe it's going to do, but you know, it's, see what happens is, you what was your, what did you think, sorry?

Sai Nelaturi

Yeah, thanks Michael. So I was working in AI, I have been working in AI for at least 15 years at this point. So when we were at PARC, we were thinking, you we were funded by the Department of Defense and DARPA and they were... thinking about how to bring lights out manufacturing to the US, especially in high mix low volume manufacturing. And one of the challenges there is to be able to, again, transfer design to manufacturing from an individual part level. How do you take a part, a CAD design, and given a set of tools, how do you make it? And fixtures, of course. Hard problem that NCEA programmers spend a lot of time solving.

Michael Finocchiaro

Right.

Sai Nelaturi

But we were seeing whether AI can actually help us there, like in 2012, 2013, right? And at PARC, we had a team that had expertise in geometric reasoning coming from my background and also in AI planning, right? So we actually built search spaces for manufacturing and we solved them, right? And we built a product and it got acquired. But the point I'm making is that AI planning or AI for manufacturing has been something that people have been looking at very closely for a long time. And for me, I think it's a very real application because you have an enumeration problem. Whenever you have a problem where you have multiple possibilities, like if you take a part and you try to figure out how to make it, there are many ways of making it. are astronomically many ways of making it, in fact.

Michael Finocchiaro

you

Sai Nelaturi

And so trying to figure out the right way of doing it is a problem that requires human input expertise, as well as the ability to be able to look through all of these combinations and provide very useful feedback. So I think of it as a human computer partnership, and we've been thinking of it that way since then. But the computer plays an important part of traversing that search space. So search is an classic AI problem. So when we solved it, we had confidence that this problem would scale, or the solution would scale, to the assembly level. And when Chat GPT came out, and lot of the research, I suppose, in machine learning came out, I was thinking about it more in the context of how it influences search. Less about LLMs as such, right? Like and more about how search and AI can be solved and what LLMs are doing in my mind is a kind of ⁓ very very high level search, right? Like because if you have a token you're generating the next token the next token How do you decide what is the next token? You've trained a model to be stochastically precise to pick a high probability next token, right?

Michael Finocchiaro

Okay.

Sai Nelaturi

And in a good search algorithm, you're always deciding what is the next action I should take. In Google Maps, for example, you go left, go right. How do you pick that direction? So training a good search algorithm with machine learning, now ⁓ to be able to pick the right heuristics, as we call it, to be able to get to the destination you want to get to ⁓ is really, think, one of the greatest areas in which machine learning can help the engineer, right? Because engineers want to see determinism.

Michael Finocchiaro

Mm-hmm.

Sai Nelaturi

They are not happy with stochasticity. They want to see, explain yourself to me. Why did you do what you did? And determinism and explainability are important AI problems. And I think the mixing the classic AI approaches of search with machine learning is the way to go. And I think it's a very real application for manufacturing.

Michael Finocchiaro

Yes. Very cool. How about you, Sarath? You were probably using it already at SpaceX before,

Sohrab

Um, so to clarify, I was never at SpaceX. I found my own rocket company, Space Ride. Uh, but so my use of my first real use of AI was years before my company. So I was an early employee of cruise automation, the self-driving car company. joined the company as employee number 10. And the biggest use case that we had at the time for AI was computer vision. Basically point cloud segmentation, helping with localization, helping with figuring out who's around us, and then using AI to predict how they are going to move around. So ⁓ the entire perception stack of autonomous vehicle ⁓ heavily relies on use of ⁓ AI, primarily computer vision again, ⁓ but also other ⁓ types of machine learning are useful to basically build models to predict how people move around. Then ⁓ less of AI was used back when I was running Space Ride. ⁓ Mostly because it's a very traditional type of a design approach and you go design and then build and test and whatnot. But then when we wrapped up Space Ride after six years and re-sync with ⁓ my co-founder Kevin, which him and I got connected many, many years ago at Cruise and worked together, you had that aha moment of what is... Parallel hardware development. What is making hardware development so much slower than software? And in my mind, there are two things. The process of software development is more straight, as in, yes, there are some iterations, there are some code reviews and changes, but more or less in one or two shots, you get to where you want. Hardware, not so very much. It's very iterative. And part of the reason is because engineers are unaware of things such as manufacturing limitations, such as availability of components, or I don't know, pricing. When you design a part, at the end of the day, there is a question, can we afford building this? Can we afford selling this or not? And if the answer is no, then have to go back to the drawing board because now you have to cheapen the design. So these are all the reasons iterations happen. But the problem begins when you start these iterations. because the process of mechanical design is very much manual. When you change one constraint, which is how geometries are related, when you change one dimension, it's not just one number that you change. Now you have to go through the entire stack of the design to see this thing that I changed, what else needs to be updated, what else needs to be updated, what else needs to be updated. And one single mistake means that you go manufacture a part, part comes back and it doesn't fit. Now you're out of your time, you're out of your money. And that is the worst thing that can happen to a So that was the moment we realized that there are a lot of relationships to be satisfied. There are a lot of constraints to be satisfied, specs to be validated, and these are done very much manually. And it is repetitive because we go through many iterations. So every iteration, you got to go and check, you got to go and check. And that is exactly what AI is good for, to make sure that these requirements are satisfied. And not only just check against, that. But when it sees violation, when it sees parts of the design that are not manufacturable, warns you and proposes an action. How can I change this to make it manufacturable? But the way we are doing it that AI proposes and doesn't go rogue and change things. allows the designer to always remain in the driving seat, allows the designer to battle that hallucination, that the exact problem that Sy said that will get explainability. Why the hell do you want to do this? It proposes, but now you're still in charge to see that does it make sense? And if it makes sense, single click. You do 100 clicks worth of action. You know it is correct. You know all relationships are preserved. So that's how we came to the approach that we are taking.

Michael Finocchiaro

Exciting. ⁓ So that makes it the other ⁓ thing that's a bit ⁓ hype is the whole idea around vibe coding and using AI for code. ⁓ I'm sure that both of you both use a lot of AI and I mean, because of cursor and well, actually last year was more of these plugins like AI Genie and stuff that were inside visual code. how are you guys, has AI changed the way you develop code? Cause you're both programmers. What's that change look like, Saurabh, in terms of just development?

Sohrab

I I primarily use ChatGPT, so I don't use ⁓ a copilot that is integrated in my IDE. I don't use ⁓ Cursor very much, so I primarily use ChatGPT. Part of the reason is I do like to be in control, as I mentioned. I do like to be the one who is deciding how the software is ⁓ being developed, the path that I'm taking. But I heavily use chat GPT to check syntax. There are a lot of basically nuances and syntaxes that I don't remember off the top of my head. And in the past, it used to be, OK, Google search, try to find it ⁓ like some talked about in Stack Overflow, and spend five minutes to find the relevant post.

Michael Finocchiaro

Ha ha.

Sohrab

Now I don't need to do that anymore. It's just one prompt check. Okay. What is wrong with the syntax and boom, that is what's wrong with that syntax. And then I go back to my ID and start developing. it's definitely been helping me, but I don't use AI in a sense of vibe coding. I am not a believer per se in vibe coding.

Michael Finocchiaro

How about you,

Sai Nelaturi

Interesting. So basically the way I do it is I use multiple models. I use cursor and I use integrated AI models inside them, inside cursor. And the way I look at it is like, you know, when I ask one of these models a question, I don't immediately ask it to write code for me. I ask it to help me think through the problem. to see if we can arrive at a spec together. So for example, if you're solving a mechanics problem where you're trying to understand how do I figure out the forces that are in equilibrium. Just a question, just a simple example to throw out there. If the AI can tell me I'm going to use the...

Michael Finocchiaro

OK.

Sai Nelaturi

basically solid mechanics to be able to figure out how to do it write the equations these are the equations I'm going to work through and I'm going to implement this and then I'm like, okay, that's fine. And then let's see the code right like and write a test and see how it works out. So I basically want to first evaluate properties like how I write, you know, how I do work in general is to see can I build a black box that gives me the input and outputs that I need that I expect before I instrument that black box, right? And if I can get confidence that it gives me that, and it usually writes some basic code to be able to give me that level of visibility. And then I go into the code and start instrumenting it line by line. Because if I can't get the output that I want, then what's the point of even trying? So ⁓ I go hierarchically that way. And when I...

Michael Finocchiaro

Right.

Sai Nelaturi

play models against each other and ask them to review each other's work, so I can quickly converge to a solution. So it definitely helps me, and especially in some fairly advanced applications involving, say, computation geometry and things like that, and on the GPU, for example.

Michael Finocchiaro

Thank

Sai Nelaturi

doing the kind of index arithmetic that you need to do on a GPU is manually very, very painful and slow, but the AI can come up with solutions pretty quickly. So we use it quite a bit in C infinity, but we're always very careful to check that the output is something that we can trust, to Saurabh's point.

Michael Finocchiaro

Yeah. Right. Very cool. Very interesting, both two very different approaches. ⁓ So when the person's using C infinity or Hestus, ⁓ I ⁓ suppose there's both AI on the surface where they're actually using it, and there's probably some AI under the covers where they can't see. Where have you actually applied AI in the solution as is exists today?

Sohrab

I can take it. So ⁓ the way we crafted our solution is that How often if someone asks you to, Hey, Michael, you've been doing this such and such stop doing what you have been doing for 40 years. Do it my way. What is the answer? I can tell you the answer is the answer is middle finger. Well, yourself. who are you telling me how to do my business? So that's what we realized that we have to meet the engineers where they are.

Michael Finocchiaro

Yeah. where they are, right?

Sohrab

Where are they? In CAD. What are they accustomed to? CAD UI. So, okay, let us build a system. Let us build a software that communicates with them in the same UI that they are accustomed to and they understand. Now, what do we do? We basically look at what they're doing.

Michael Finocchiaro

Right.

Sohrab

predict ahead what are the steps that they can be doing exactly that search concept that ⁓ Saheb brought it up. There are different paths that you can take. And then try to even refine the prediction. What is the most likely action that this designer, this engineer is going to take for their next step? And propose that in a graphical overlay. with icons that they are used to that, okay, this is what I'm proposing to do. And if you agree, press enter and move on. And if you don't agree, just draw whatever the hell you wanna draw and it goes away as if it never existed. So AI is the part that predicts what are the options. ranks the options and over time within the session learns from your selections, learns from your actions, tries to figure out what are the areas that you need more help and what are the areas you seem to be fine on your own to refine that proposal, to refine that best action that is going to go next. So AI for us is all under the hood. I don't change how you work. Visually, I don't change how you basically interact because I think that is, I don't know. I think it's fool's errand trying to change engineers who have been doing things the same way for 20 or 30 years. It's probably not gonna happen.

Michael Finocchiaro

All right. Interesting. had the same approach or you have a completely different way of doing it?

Sai Nelaturi

I really like what Sarab was saying. So we think similarly at C -Infinity, the way I go to my customers and try to ⁓ help them is to say, let me meet you where you're at today. ⁓ But let me also try to show you where you could be and let me take you there. And so when we talk about AI, everybody wants to understand AI and see what they can do with it. And interestingly, a lot of the large corporations that we speak to, have a mandate to go and pursue artificial intelligence. And as a result, they're trying to understand the various types of AI that exist today and how they can truly benefit from it. And they have questions about determinism and all of these things we talked about. proposing an entirely new interface is a non-starter to Saurabh's point.

Michael Finocchiaro

Yeah, of course.

Sai Nelaturi

⁓ You want to minimize friction in existing workflows, especially that CAD is just one part of the picture, right? Because you very well know, it's, you have PLM systems and you have ERP systems and like you, and design, is no, know, every design is a redesign. You never ever. Restart with a blank slate unless like you have a completely new manufacturing process, know Typically manufacturing is ahead of design you if you has a new process that you come out then you try to design for that process You don't design something in an inventor process to make it right like it's that's a rare occurrence if at all So my point is because there's ⁓ all of this Redesign, you know and all of it is connected and stored in PLM systems

Michael Finocchiaro

Alright.

Sai Nelaturi

you have to be able to harmonize CAD, PLM, and all of these things. So there's an integration layer that has to be like something that a customer is convinced by, right? And that you can actually get all that information together. And then under the hood, you want to be able to show powerful time-saving capabilities, right? Like where the AI is really doing the things that people expect to do ⁓ or do manual work, for example. In our case, it's more like planning and documentation. So if people are trying to figure out how to assemble a product, then they're thinking about, I've done it this way in this factory before, so I want to reuse all of these components, but here are some new parts for this new ECN, this new revision that came through. How do I fit it in? They have to explore a bunch of configurations for different products and figure out how to do it right. So that takes time. Now, if you could do all of that retrieval search, and ⁓ generation of instructions and plans very quickly, the user can interact with them and okay it. That's a huge time saver. It's a $50 million year over year problem for large OEMs, right? So that's how we see the benefits of AI directly in terms of operational efficiency, as opposed to some futuristic evaluation of...

Michael Finocchiaro

Huge.

Sai Nelaturi

new design concepts with a new interface. We like to get there. mean, our mission statement is to design the designer, but we want to go there from meeting people where they're at today.

Michael Finocchiaro

Makes me wonder also about ⁓ the models. Are you guys training your own models that then understand precisely, ⁓ in your case, has SORAB more the design for manufacturing? maybe I'm saying wrong, maybe there's an agent and it really understands these kinds of holes and this kind of material creates this kind of stress or whatever. And Sai, it's more the contact surfaces, right, when you're doing the assemblies. that you're working on. Have you guys, are you training your own models? Are you just inside your software or are you leaning on the open source or the Olamma versions of them? And I'm a question is like, if a user puts in a prompt, does that also go in or is that like a separate, there's like a, you're training on predictive responses. the last time the user said this, he wanted this next.

Sohrab

⁓

Michael Finocchiaro

But I'm not going to believe everything he says because something he said in this thing may be false. And I've got this other knowledge base which has the objective truth around the design. Sorry, I didn't want to interrupt, but I wanted to make sure we cover that.

Sohrab

So for us, we do design our own model. It's a completely brand new model built from ground up. ⁓ One purpose to help mechanical design from the basic beginning of sketching, then later on to 3D, later on to basically ⁓ assemblies and GDNT. And then throughout every step. have manufacturing in mind. The reason we are doing that is because by focusing on building a custom model that is really trained at being good and accurate at these problems, we can have a very small, very nimble model that is fast to run. but we don't basically rely on the other parts that, or we don't need the other parts of these LLMs that we don't use. We are not building general smartness. We are building intelligence specifically for one very specific niche application. And that's why we are doing that. The other reason we are doing that, is a lot of, as I'm sure Sy also has experienced this, there's a lot of spatial reasoning, geometrical reasoning and understanding that is required to do what we do. And LLMs are not trained for that in particular. And when people put them into tests, a lot of times LLMs were not trained for certain things and happened to be good at it. Geometrical reasoning, spatial understanding was not one of those things. One of the recent comparisons that I've seen, someone showed pictures of different rooms to different elements and asked them to, now seeing this, put together a blueprint. At best, it performed at random, far worse than how human does. And I'm sure you can train a tiny model that is specifically designed to solve that problem and it can excel. That's why we are building a custom model. It also has its other advantages that, example, the inference cost is going to be not a problem. A lot of these companies who are wrappers around chat GPT or Anthropic or whatnot are bleeding cash with every API call. ⁓ so you're at the mercy of them. ⁓ So that also solves that part of the problem from basically having an awesome margin. But the main reason is we want a model to do a specific

Michael Finocchiaro

Yeah.

Sohrab

and do that really well.

Michael Finocchiaro

So the agentic approach, like just really nail the one thing perfectly rather than, ⁓ okay, very cool. How about you,

Sai Nelaturi

Yeah, I mean, we are building our own models as well, right? Like, so our models are, as I was saying earlier, a mixture of search and machine learning. So we call them neuro-symbolic, right? And because it's a mixture of the neural parts of machine learning and the symbolic architectures you have in traditional AI, because we want to be deterministic and explainable, right? Like, you can't get that from off the shelf LLM. So... So the, but at the same time, I always believe in hybrid ⁓ representations, right? Like they usually tend to bring out the best of both worlds. We've seen that in CAD over the years, right? Like, and it's just the evolution of things, right? You start from distinct places and then you merge. So the way I see it merging eventually with ⁓ what the benefits of LLMs are. is when we eventually have to convert all of that spec that I was talking about and getting the instructions and all of that on all the planning done through AI, but then writing it up for people to be able to ⁓ read eventually if the human needs to be in the loop. In our case, we also envision a path robotic assembly because we're talking about being compiler for physical AI. So eventually you want to get all the way to instructions that ... that a robot can go execute, but if a human needs to read the instructions, right, like then, you know, we can generate that. So I think that's trying to get the best of both worlds at the right time. I don't think foundation models for engineering as such, right, like are, I think it's a good goal to work towards, right? And so I'm glad people are thinking about it. But I like to start with specific models and specific problems that we are building ourselves as opposed to solving the broader problem upfront.

Michael Finocchiaro

Interesting.

Sohrab

I very much agree with what Sai said. I think LLMs have a role to play here and maybe that role could be, so you mentioned the last bit of the thing that, okay, all those steps documented so a human can read and understand. There's also the beginning part that maybe, for example, you want to start the design and you want to lay out the specs. ⁓ I want to manufacture this using CNC or maybe injection.

Sai Nelaturi

Yeah, the requirements,

Sohrab

this needs to be IP68 rated, blah, blah. So that basically layer to take specs and requirements from you and then under the hood, translate it to something that is now enforceable in form of constraints, in form of relationship, in form of geometry in CAD. And then when you start designing, then it switches to that model that is okay. Now I understand these are the ⁓ requirements that I have in terms of geometrical relationship, geometrical requirements. And as the person is designing, let me use that specially trained model to help them with that.

Michael Finocchiaro

So I'm thinking too, like, it's interesting, because I saw a post today from a woman named Leslie Gao, ⁓ who was talking almost like, it sounded like I was reading myself, because she was talking about the history of CAD kernels and why we haven't had a chat GPT in the CAD world, right? We haven't come up with that. she was talking, and it's true, there's a lack of fluency. in terms of geometry, right? By the deal, the elements are trained on Reddit. They're not trained on Katia and ProE and Creo and that stuff. So do you guys see like... You know, how does what's the roadmap for that? Like I talked to Leo, I talked to Mauer of Leo AI yesterday. He's been in the news because he got a big raise over the summer and he actually built a lot. He called an LMM a large, large mechanical model. So he tried to put all this knowledge about mechanical engineering directly in so that his LMM understands the best practices and stuff. Is that sort of the vision you guys see too? Do think we'll get to the point where there'll be a a generic ⁓ LMM that you guys could also draw on and you wouldn't have to reinvent the wheel by retraining your own model with all that stuff? Or what do you think that looks like two, three years from now when if AIG doesn't come and we're all wiped out by Terminator 2, of course? ⁓

Sai Nelaturi

Well, mean, can I take this one? Okay. You know, I wasn't sure where we were going, but, so, okay. I think that, I think that, ⁓ you know, there's a lot of richness in how design intent and manufacturing is expressed, right? Like even if you're doing, if you're, if you're taking an a single part and trying to model it, have a feature tree. You have all the modeling operations that you have inside CAD, right? Like that need to get stored. Design into, there are design features that are manufacturing features. How, you know. ⁓

Michael Finocchiaro

Right.

Sai Nelaturi

CNC programmers going to interpret the features are going to be maybe different from the designers in turn. What's your fillet radius? How does it correspond to what you have in your organization, right? That kind of information is going to be impossible to train on because it is going to be proprietary, right? Like and it is going to be, you know, very hard to generalize because the representation isn't straightforward. It isn't there isn't a large corpus of text, you know, the equivalent of text to be able to at internet scale effectively. to be able to train large models. So the best approach to even try to attempt such a thing is to create synthetic data. So you have to run your own simulations and craft your own outcomes and train on it. But it's by design, ⁓ only a first step, is a first approximation. And it requires fine tuning inside deployed systems, inside enterprises. So I think that the... Capturing all of that intent, know, the actual practice of engineering is going to be very hard because there's no equivalent of Stack Overflow or Reddx for, you know, mechanical design.

Michael Finocchiaro

Yeah Well, maybe the comment feeds on a G2 or something, right? Not even. Sorry, go ahead, Sarab.

Sohrab

I could not. No, no, sorry to interrupt. I often feel like a broken record keeps saying the same thing and people not understanding it. I'm really glad. I would have hugged you if you were here in person. Like, what out there is 100 % correct. Now let me give you an example. The first rocket engine that we designed at Space Drive to save space.

Michael Finocchiaro

Hahaha!

Sai Nelaturi

Exactly, yeah.

Sohrab

We assume that, okay, we are going, and also manufacturing costs, we are gonna get a standard pipe. We are going to bore the inside, a little bit of lip at the beginning and the end, and then machine these huge chunks of aluminum that are gonna slide in and then bolt radially like that. Amazing design. Engineers showed it to me. I was blown out of my mind. What's wrong with that? What do you think?

Sai Nelaturi

Zoom accessibility.

Sohrab

Try to assemble that. I want to kill you. Imagine lining up this chunk of aluminum that weighs about easily, easily 80 pounds to precisely going and you have 52 radial bolts. So even slight offsetting angle can make it impossible. Now you've got to torque this shit while it's inside that. Where is that?

Sai Nelaturi

Yeah.

Michael Finocchiaro

Yeah. you You're host, right? Not happening.

Sai Nelaturi

how you're gonna hold it up. need a, yeah.

Michael Finocchiaro

Yeah.

Sohrab

so every, exactly, so, ⁓ how are we gonna hold this to line it up? Okay, we need this, I don't know, ⁓ moving cart. Now we need something to basically push it forward while it is steady. Okay, let's design this other thing. Where is that information recorded? Nowhere. in the designer's head that they wanted to kill themselves from the fuck up that they have done and they guarantee that they're not going to do it again. There is no information like that anywhere in the world for you to train on. And that is the, so I actually use all of these text to CAD models and I asked one question, sorry, two questions. Number one question, design me a bookend. Okay, they can all design a bookend.

Sai Nelaturi

and we'll...

Sohrab

And then the next question, design a fuel injector for a rocket engine. And how it does, it gives me a summary that probably I could have gotten from Wikipedia. And that's the end of it. OK. So here's the thing, if I'm designing a ⁓ rocket fuel injector and I don't have that basic information, then I have no place designing that because sure as hell I'm going to blow up myself and the entire test path and no amount of regurgitated Wikipedia and whatnot is going to save me from that. So that's, it would be amazing if one day we have

Michael Finocchiaro

Right.

Sohrab

Jarvis for real and said, okay, design me a spacesuit such and such and said, look, here's the spacesuit. It be amazing. Don't get me wrong. I wish we could have that. Yeah, but definitely not that soon.

Sai Nelaturi

And we should work towards that.

Michael Finocchiaro

I well on Wednesday, this week I had a, there was a Scenera, a simulation startup from Germany and they had a nice conference called Engineering the Future and I had a talk about the future of engineering and I talked about Elon Musk, who's not my favorite, no I'm sorry, Jensen Wong talking about the factory of the future, the physical factory will always have an AI factory. And I said, well, what we're doing collectively, the three of us here and all the other people in this industry is creating that AI factory, right? We're slowly building that factory that will eventually replicate itself and build stuff by itself. I think what we end up is because of everything you just said, you end up with these relatively small, very limited, almost deterministic agents that could do the one thing and then you accumulate up from there where I think of it like. I want to have an amenities agency. like, I don't know about that. You know, are you really sure? we recalculate? You know, there's always be one guy trained to like doubt instead of you saying how wonderful you are and every idea you have is the greatest idea since I spread. You would say, I don't know. That's just okay. I think you got to work on that some more.

Sai Nelaturi

You know, I like the analogy to an AI factory because ⁓ building on top of our earlier discussion about specialized models, Just like in a factory, an assembly factory, you have specialized cells, you know, to do very, you know, specific activities because you're either like, you ⁓ you have, need the right tools, the right fixtures, and you need to have the robotic systems, the gantries and all of that to be able to do things. You can't just instrument and, you know, reconfigure a factory at will.

Michael Finocchiaro

Right.

Sai Nelaturi

you know, because you have all these specialized skills. And imagine in an AI factory, if each of these specialized skills were specialized models, mini factories in some sense, right? So Saurabh's model talks to my model, and then my model gives him feedback, he gives me feedback, and then we collectively coordinate and then ⁓ give output, right? That is an agentic view of AI factory that potentially can actually be successful, right? And I think that, but the...

Michael Finocchiaro

Yeah.

Sai Nelaturi

you know, building process is bottom up. You can't start from saying, I'm going to build a factory. You're going to say, I'm going to build the specialized things and then put them together in a way that they can actually talk to each other. Interoperability becomes a big part of it too, of course.

Michael Finocchiaro

No, you gotta go.

Sohrab

⁓ I think one of the things ⁓ to Sy's point, ⁓ you need to understand how things are done. So there is this interview that Gary Tan of ⁓ WIC has had recently and he said, you need to have deep empathy for how the user of what you are developing for does things. If you have no understanding of how they go about their way on a daily basis, how are you going to make their lives easier? If, yeah. LLMs, chat GPTs of the world, these models that you can basically communicate and they generate videos and images and whatnot, they're cool. No argument about that. I I use it with my son to generate funny graphics every day. It's just him to laugh at it. But the question is, Is that the problem that that mechanical engineer has? Is that the problem that technician at the factory on the factory floor experiences? Or is it a different thing that they need help with? And if I've never spent a day designing a component, if I've never spent a day machining a part, I spent an entire summer working as an assembly person on a gearbox manufacturing facility. My job, for example, one day was to literally bang in this shit in and then the next day move to another location. If you've never experienced that, how are you going to develop a solution to help that person?

Michael Finocchiaro

Yeah. Hmm. Great point. ⁓ And I think makes me, I'm going to try to come up with, remember a question I have for the end because I've got it. I think I got a good one. ⁓ If we look at then the customers of ⁓ C infinity and Hestus, like I always think of digital maturity as being a sort of a spectrum from one being, they're still using email and Excel to do most of their collaboration to five where we're talking about the AI factory. does, it's all completely agentic and adaptive. And nobody really is at five, right? If we're honest, everybody's somewhere between one and two or maybe, two and a half, probably not very many, three. ⁓ but maybe that's maybe my perception of the world. You guys are in California, you know, everything's much more beautiful and shiny and stuff. ⁓ where is it? What's your perception? Obviously don't name the names of the customers, but when you're going to your customers, do you get an idea of where they are on this maturity with or without AI, right? The AI obviously is that four or five, but Where the customers before you come in, where are they? then when you guys put in these awesome solutions, is there sort of an aha moment where the customer's like, well, if we actually fixed data governance, we didn't have the data siloed across each department. And the MEs actually talked to the guys in simulation and manufacturing, and there's a little bit more people talking to each other. Maybe things would be better. So I'll let you guys answer that one a little. ⁓ Take your time.

Sai Nelaturi

I think that when we go to customers from my perspective, we lead with the problem that needs to be solved. AI is a solution or an approach to solve a problem. If it's iterations, they're sort of upset. ⁓ Or if it is just the pain of doing day-to-day process planning, design for manufacturing checks and all of that, that's a problem that people have. on a day-to-day basis, talking about empathy with your users, right? And when you lead with a problem and say, can help you solve this problem faster because AI gives us some capabilities to do it, the reception is just a little better than saying, let me give you some AI, right? And are you ready for it? And that's one part of it. So leading with the problem is crucial. and an understanding of the problem. Like we were talking about how it all connects with PLM and CAD together. ⁓ That's critical. But the customers that we've spoken to, as I was saying earlier, they have a mandate for AI and they want to see how it can help. So they're receptive. There's a spectrum. There's a spectrum. I've seen in ⁓ top enterprises that there are some really visionary people who are

Michael Finocchiaro

do get a feeling of where they are today? Are they closer to one? Okay.

Sai Nelaturi

who are pushing forward with custom-built language models or custom fine-tuned models and ⁓ really creating an agenda and a roadmap for how they want and a vision that they can execute for AI. So I've seen that. And I've seen people also saying, like, I don't know what to do with this necessarily. Everything is scattered in terms of how I organize my work. Can you help me solve this very, very, very specific problem? right, like without talking about the vision as such. So there's a spectrum, but there are people I think who are at one end who are thinking like, I think this is going to be a game changer and I have an understanding of how it affects my enterprise and I have a plan for it, you know, and there are other folks who are trying to get, ⁓ get onto it. They get so as you would expect, I think.

Michael Finocchiaro

Mm-hmm. What about in your experience, Zorab?

Sohrab

So I guess you can say the typical mechanical engineer is probably a little bit higher than one or two because they have to use a lot of digital tools for design, for simulation, ⁓ to be able to do the routine job that they do. So they're a little bit more advanced than email, even without the new world of AI. But... ⁓ I've seen similar things as Sai mentioned. So there are some ⁓ enterprises and it's mostly enterprises that have AI policy. And I think when it comes from management down, I don't think it is going to be successful. think it is in some cases, it is driven by a visionary person, as he said, but in many cases it is a way to, okay. It's a, ⁓ I don't know, bandwagon that if I felt behind from it, my stock value is going to go down because people think that I'm not adopting this new, beautiful thing. Like I heard a while ago was Bumble CEO was saying that, ⁓ in the future, AI would have a role in like the AI of this person and AI of that person will talk and see if like.

Michael Finocchiaro

Hmm.

Sohrab

We are talking about one of the most basic human connections and we want to use AI for that. Why? So the way I look at it is exactly as Sai said. Users have issues. Users have problems. What is the problem? You ask them, ⁓ I have to do this thing. It's repetitive. It's time consuming. It's not creative. I don't enjoy doing that. I wish there was some solution that would automate it for me. To them, it doesn't matter if it is AI, if it is a genie in a bottle, or if, I don't know, it's just whatever the bit. Maybe you hire someone and put it in a dungeon and they do the work for you. As long as they don't have to do that work and it happens correctly and efficiently, they're satisfied with it. And that's exactly, I think, how we should look at AI. What are the problems?

Michael Finocchiaro

you

Sohrab

Can they be solved by AI? I claim there are problems that cannot be solved by AI. The fact that someone needs to find a companion, I don't think that's a problem for AI. There's a lot of chemistry nonsense involved that I don't think AI is designed for that. But here's the thing. There's a lot of repetitive steps in assigning constraints and dimensions and whatnot, and that's something that is trainable and there's data for it. Yes. AI is good for that. So let's use it to solve this part of the problem for them to have more time for stuff that they can do well. That's my approach. I want to say there are three, there are very receptive to solutions when you approach it like that. Tell me what your issue is. Let me show you a solution. And if you like the solution, you can adopt it.

Sai Nelaturi

Thanks.

Michael Finocchiaro

But when they, the customers when they're using has to see infinity, are they, do they have that feeling like, wow, if I was better at doing data governance, if I was better at collaborating between departments, or is it sort of just a slow ripple effect? It's just part of the whole slow march towards ⁓ progress.

Sai Nelaturi

I yeah, go ahead, Saurav. I think that's a mixture of both, right? Because there are bandwidth issues or impedance mismatches between different siloed departments, right? And that's obviously something that can be streamlined with better communication. Communication is obviously a big problem that we can actually solve to some extent with AI. And the

Sohrab

I mean, okay.

Michael Finocchiaro

Right.

Sai Nelaturi

But the other part of it is that the problem itself is challenging and repetitive and tedious. In its nature, it needs to be addressed. A new solution needs to come out for it. Otherwise, we're not going to make progress from where we're at today. Data governance and data harmonization are enterprise-level important problems. I think that needs to...

Michael Finocchiaro

Alright.

Sai Nelaturi

be solved because that obviously creates friction in efficiency. So there's that level of that layer of problem, which I think we absolutely need to address in a full solution, like in a product that we deploy. ⁓ But I think there's two parts to it. There's that communication aspect and the actual problem itself that we need to solve. So that's how I view it.

Michael Finocchiaro

Okay. How about you, Sarab?

Sohrab

So I mean the nature of tools that we have right now is that it is integrated within CAD. So when they leave the CAD right now we don't work with that. So that's obviously an area data governance that you don't get involved. And because we don't take the data out of cloud or have it imported into CAD, we don't have to necessarily deal with that issue right now. But as we go into later on design for manufacturing and other aspects, obviously being able to seamlessly hand off from one software, say CAD, to another, to another.

Michael Finocchiaro

Right.

Sohrab

would be extremely valuable. I mean, even today, for example, when people design and then they go analyze and say you design in, I don't know, fusion, solidworks, whatever. And then you want to go to ANSYS to, for example, do analysis. There is a lot of model cleanup and steps that. Yeah.

Michael Finocchiaro

Right. Or to go to Gibbs cam or Mastercam to actually manufacture. That's another set of cleanups.

Sohrab

There's definitely room for improvement there to say, okay, here's your model. whatever that is, and you want to take it to Ansys, you want to take it to ⁓ any other analysis tool, maybe you use CommSol or whatever, let me help you with that. I think that is definitely an area that is ⁓ prime for improvement and people would appreciate that you open to solution. ⁓ This is not an area that you're currently focused on. It's definitely something that we would be looking to, but not an immediate thing. And that's why we haven't

Michael Finocchiaro

Mm.

Sohrab

and had any ⁓ data governance issues up until now.

Michael Finocchiaro

There was one question from ⁓ Uzair asking, what's the craziest thing you saw AI do in MCAT? I'm not sure how to answer that one. And then he also said your thoughts, if you had any thoughts on Autodesk releasing their own foundational model for Revit and Fusion. I hadn't even seen that they had done that for Revit. I've seen it for Fusion from AU.

Sohrab

⁓ So there two things that I can say. ⁓ I think any... Example that I've seen so far from models being generated in CAD, the most sophisticated one or maybe the most common one that everyone goes to is, design me an Arduino box, design me a Raspberry Pi case. Don't get me wrong, it is nice, it is creative, but then I ask myself, why would you need to do that?

Michael Finocchiaro

Right. Ha ha.

Sohrab

That shit is 10 bucks off Amazon. What a stupid thing. I understand that you need to take stepping, like little steps to get to where we want to get to. But this is the way I look at it. One of the biggest problems of CAD is that there is no CI CD. There is no, okay, I changed this part of the code. Let's run all the tests on it to see if I have or have not broken other parts of the code. There is no such check. So if AI gives me a solution, oh, here is the entire geometry design for you, and say there are 200 bolts in it, I have to manually go through every single one to verify if there is a mistake in one. And maybe there is a mistake in one, and that is less than 1%. That's half a percent. That is still going to screw my part. So even if you are 99.5 % accurate, you're still going to screw me.

Michael Finocchiaro

Right. Hmm.

Sohrab

And I still have to spend equal amount of time to go through. That's why I think AI incorporated in CAD and generating models, like 3D models, that's what I'm talking. is not the solution. The solution is hand in hand with the engineer, step by step. You do one thing, the engineer does one thing, AI proposes next step. And then when they accept it, they see a chunk of work that is visually accessible, that they can see that, okay, in a fraction of a second and maybe a few seconds that, yes, that is good, accepted, approved, let's proceed. Let's move on, let's move on. So it doesn't do correctly the work for you, but now you have to spend equal amount of time that you would have done designing it, reviewing it. How good does it make it?

Michael Finocchiaro

Okay, we're almost at the top of the hour. had one other question and it's more for the maybe some of the younger listeners we might have. Like, you there's a lot of AI generated anxiety over is AI gonna take our jobs? I think from what both you guys said already, the answer is no way because there's always gonna be some human stuff. But if you were to address the people that are studying engineering and they're gonna go into the job market and are thinking, what do I need to focus on? Because I get this question actually, through LinkedIn because people read my post and they're like, if you know, do I, you what do I do? Do you guys have some advice to those people? Like what to focus on? What are the areas that will never be delegated to some crazy AI where they'll won't be out of a job tomorrow? I mean, I've had, I just heard three people got laid off yesterday, friends of mine, because AI replaced their jobs. So do you guys have any, any perception on that?

Sai Nelaturi

mean, since the industrial revolution, right, like there's always been this advancement of technology and the need for people to keep up with it. And whether it is computer programming or the internet or whatever, right, like major paradigm shifts that have happened over time, people have, we've had the same conversation, right, like, and you, you know, it's, we adapt, we survive and we evolve, right, like, and.

Michael Finocchiaro

Yep.

Sai Nelaturi

⁓ So I think that it's about learning skills and thinking about how you do work. mean, like, you know, when Wikipedia came out, people stopped going to the library as much, right? And so, you know, when you actually have a new way of organizing your information, you know, or Google, right, like you start changing the way you think and changing the way you work. Now what we have is a new set of tools, basically, and that's going to influence the way we're going to think and work. If we think about it that way as a skill that we need to learn, I think it's easier to sort of manage it and adapt to it. ⁓ But at the same time, I can totally understand if that, you you're seeing the power of these tools being able to replace some types of jobs, then ⁓ that is a real concern. And the concern can't be alleviated other than embracing it, I think, because if you're fighting it at this point. it's going to be a very hard battle to win, because there's a lot of investment in artificial intelligence. The world is kind of moving in this direction, And so the best way to do it is to figure out how you can learn from it and adapt, I think. And it will be fine until the next paradigm shift, you know.

Michael Finocchiaro

Do you agree, sir? Hahaha

Sohrab

I think I agree. So AI is definitely going to create a paradigm shift, like many other things that we have seen in the past. So textile revolution, industrial revolution began with textile machinery. And to be honest, those people who oppose it, I do not know how they felt about it. But this is one thing that we know. Do we have as many jobs as we had before working in textile, probably fewer people would have bigger output today, but still a thriving industry that is employing a lot of people. The other way that I look at it, when email became a thing, people thought that that's the end for postal services. The same technology brought us Amazon and online shopping. And I claim that delivery services are having a prime of their time, like packages after packages being delivered to apartments and houses and whatnot. So yes, there is going to be disruption. Yes, there are going to be some changes in how we do things. And some jobs might reduce in number, but then we are going to introduce other jobs because of it. Or that's what I think.

Michael Finocchiaro

Mm. new jobs. ⁓ That's great. I like that. And I appreciate you guys both. It was really insightful, very helpful. I hope you guys had a good time too.

Sai Nelaturi

conversation. Thank you, Michael. Nice to speak with you.

Sohrab

Same. I I'm glad that because of this, and I got to know each other and I'm definitely going to have a coffee or whatever you want at some point.

Sai Nelaturi

Exactly.

Michael Finocchiaro

I gotta come out to San Francisco and have a coffee with you guys too, right? Yeah, there you go. When I do my famous startup conference, you guys will be stars of that. All right. Well, thank you very much. Thank you everybody for joining and we'll see you on the next episode next week. I think we've got ⁓ In-Use and Spare Part 3D 2 French, French Dutch startups. Thank you very much. See you guys next time.

Sai Nelaturi

You should. Or we come to Paris, right?

Sohrab

Yes, yes, yes. Sai Nelaturi (1:00:04) soon. Thank you very much, Michael. Bye, Sorok. Sohrab (1:00:20) Thank you. Michael Finocchiaro (1:00:21) I'm just going to stop the recording.