The Infrastructure Layer: AI for Product Complexity — with Cosmon and Metafold

Episode Summary

The episode delves into the intersection of artificial intelligence (AI) and product complexity management, focusing on how advanced mathematical analysis can enhance manufacturing processes through geometry-based information conversion. The guests are Elissa Ross from Metafold and Rui Aguiar from Cosmon. Metafold specializes in converting complex 3D geometries into meaningful information for manufacturers, leveraging true mathematical analysis of shapes to integrate with AI technologies. Cosmon, on the other hand, develops agentic tools that streamline design-to-production processes using novel deep learning methods tailored to specific engineering tasks.

Two critical insights emerged from the discussion: first, the unique challenges and opportunities in applying AI to 3D geometry versus text data, highlighting the complexity of 3D shapes and the lack of analogous models. Second, the importance of strategic focus for startups, emphasizing that success lies in excelling at a few specific areas rather than attempting to dominate all aspects of engineering.

For PLM and engineering professionals, the key takeaway is recognizing the potential of AI-driven solutions like those offered by Metafold and Cosmon to optimize complex product development processes. By focusing on specialized tasks and leveraging advanced mathematical techniques, these tools can significantly enhance decision-making and streamline operations in manufacturing environments.

---

Full Transcript

Michael Finocchiaro

And we're live. ⁓ Hello, everybody. This is Michael Fenicara on the ⁓ AI Across the Product Lifecycle podcast. ⁓ I'm joined today by Elissa Ross of Metafold and Rui Aguiar, if I said that right, of Cosmon. ⁓ Very exciting startups in the engineering and simulation space, and I'm really happy to have you both. Elissa do you want to start telling us about yourself and about ⁓ Metafold?

Rui Aguiar

That is correct.

Elissa

Yeah, absolutely. First of all, thank you for the invitation. It's certainly a pleasure to be here. ⁓ So I'm the CEO of Metafold, also a mathematician by background, so I have a PhD in geometry. And at Metafold, we are really all about geometry and in particular, converting geometry into information that is helpful for manufacturers that is meaningful for critical decisions ⁓ in manufacturing. And I'm super excited to get into the intersection of all these topics with AI, but I'll just give a brief preview that our company is really based on, I won't call it traditional, but I will say kind of a true mathematical analysis of shapes. And there's a very interesting interface between that and what's going on, of course, right now in AI, which we'll surely get into.

Michael Finocchiaro

That's awesome. Tell us a little bit about Rui and Cosmon.

Rui Aguiar

Yeah, so my name is Rui. I'm founder and CEO of Cosmon, my background. ⁓ double degrees in hardware engineering, systems engineering, grad degree in artificial intelligence. What we do is, and it caused one, is we build agentic tooling across the design to production life cycles. So whenever you build something, there's 50,000 things you got to do across CAD tooling, simulation tooling, PLM tooling. We build agents to make those processes better, faster, easier, more streamlined for organizations. Specifically, we focus a lot on sort of more novel deep learning methods, agentic architecture is sort of like model development. performance across domain specific tasks in engineering that are hyper specialized without having to use a lot of IP or anything else like that that can be kind of bolted on out of the box. And we especially excel in specific design and simulation use cases as our sort of initial adopter.

Michael Finocchiaro

Awesome. It's kind of a bit of international today too, right? Because I'm coming in from Paris, so Rui, you're in San Francisco, and Elissa you're up in Canada. So that's kind of lot of fun. ⁓ That's exciting. like for, ⁓ I used to start an icebreaker question around ⁓ when the OpenAI moment happened, when the moment pivoted around LLMs.

Elissa

Correct.

Michael Finocchiaro

Melissa, were you like super bullish? Like, my God, this is going to change everything? Or were you more wait and see, a little more skeptical about, you know, how it was going to change how we do engineering and how we do manufacturing.

Elissa

Yeah, it's an interesting question. think, you know, like so many people, I had this reaction to it where it feels just like simply magical at first. The first encounter with an LLM was, you know, kind of mind boggling. But I think for me always, I come back to how does this work for geometry? How does this work for 3D data? So this has been the central question in our analysis of AI and

Michael Finocchiaro

Mmm.

Elissa

And the fundamental truth is that what's happened over the last few years in terms of LLMs and text is very, very different from progress on 3D data and geometry. And so there's no analogous model for 3D geometry. Lots of reasons for that. ⁓ 3D geometry is just so much more complex, more heterogeneous.

Michael Finocchiaro

Right.

Elissa

There are no data sets like the huge amount of text that we have. And just comparing geometry to text, text is so simple. Even two dimensional images are far, far more simple than the three dimensional case. So ⁓ I think, you know, my reaction to the language models was kind of distinct from my reaction to anything in 3D. And I think what has been kind of ⁓ maybe confusing for the market is that people expect that you can do the same things for geometry, for engineering, for anything you want actually, as what is happening with the language models. And this is really not true. have some progress, we have sophisticated things happening in other kinds of data, but it's not the same as for text. So I think that's kind of my key reaction.

Michael Finocchiaro

Hmm.

Elissa

This is not going to transform engineering the way that people might at first think it will.

Michael Finocchiaro

Interesting take, I like that. How about for you, Rui?

Rui Aguiar

Yeah, so I come a little bit more from the traditional AI world, living in San Francisco and so on. actually my first interaction was C53.5.

Michael Finocchiaro

Ha ha.

Rui Aguiar

GPT-2, GPT-3 has been coming around since 2018. Really, think what the actual, I mean, there was some innovation around, of course, the training data in the corpus and the construction of these models, but the real innovation here, I actually think was on user experience lens. I think it was in the ability to really have this thing that you could interact with in an easy, accessible way that was simple, that was sort of allowed it to take off from a virus.

Michael Finocchiaro

Hahaha

Rui Aguiar

perspective. I think it definitely wasn't clear as of early 2023 or whenever, even 2024, sort of the industrial and corporate use cases. know, it's chat bot, it's like, okay, there's pretty good consumer product at this point in time. You know, a lot of the kind of core use cases with stuff like Anthropic hadn't really been kind of bought out. I don't even think coding was even considered at this point in time as a major use case. ⁓ And so, you know, it's going to be interesting to see how these things evolve. I think that

Michael Finocchiaro

Hmm.

Rui Aguiar

In various domains, you get various spectrums of ⁓ autonomy and agentification, or whatever you would like to call it, across engineering, across legal, across doctors, across coding. You have some that take off faster than others. Obviously, coding has basically completely gotten turned around in the past three years, where now the entire field has been basically just taken over by agents. ⁓ have slower at

Michael Finocchiaro

We'll be talking about that in a minute.

Rui Aguiar

Yeah, you have slower adoption on sort of the middle grounds, know, legal doctors where, yeah, you can still create a lot of value in building out these things and then deploying it across these. Engineering is a little bit of a weird one. ⁓ you know, engineering is something that I think we've seen basically no penetration into so far, again, for a variety of reasons, ⁓ technical being some of them, probably IP and security being some of them, ⁓ skepticism, maybe being another one of them having the right use cases. What we strive to do at Cosmo is we try to strive to strike a balance, right? You get a spectrum on agentic development, you get a spectrum on agentic design all the way from like, you know, there are a lot of companies today that are like zero shot me like a car. this is completely worthless. Like, no one's ever going go and trust this thing, right? All the way back down to like, you know, more simple things of just like, hey, you know, help maybe review my drawing or help do this sort of 2D work or something else. And really, like, when you think about building out a useful agent for an engineer, it ends up kind of somewhere in the middle here. And that's still being defined. I think that's personally very exciting. I think that, like, you know, a lot of people don't know what's possible. Like, there's a lot of things we're trying out internally. But this is kind of where I see the state of the art today for the field.

Michael Finocchiaro

But it's, yeah, maybe when you mentioned the text interface, it's kind of funny too. I think that, you know, Google changed the world by just giving us a line in the middle of the screen. And then we spent 20 years building all this stuff all over the place and then chat GPT changed the line, the world by putting a line in the middle of the screen. You know, with all this other UI stuff, it's kind of crazy that the one line is the ⁓ thing. and you, you,

Rui Aguiar

Best products are simple. ⁓

Michael Finocchiaro

You alluded to the changes in R &D. I wanted to talk a bit about that, like how ⁓ at Metafold and at Cosmon, ⁓ you guys are using AI in coding or in program development. Certainly probably your PRDs are probably already done more or less by AI. ⁓ a question I came up with a week or two ago is, you when are we going to start having agents showing up for the Scrum meetings instead of, you know, in place of, or in addition to the people? So Rui, don't you pick this one first? How are you using Cosmon? How are you using AI, ⁓ whether it's cloud code or any gravity or whatever, probably not lovable, ⁓ but ⁓ how are you using it to build Cosmon?

Rui Aguiar

Thank you much. Yeah, so when you think about AI for software, the real cost in software is not actually like the lovables of the world, the zero-shutting. It's like maintenance, it's technical debt, it's recurring costs and use cases. So we try to really deploy it for like the cost centers of the business. This is, think, why something like a cloud code is so valuable. So we use that pretty extensively across the engineering team. think pretty much everyone's on it. I don't think anyone's on any gravity. In addition to that, we use some for customer support, answering use cases. And then we also use some for just like drafting, like slide generation, more classical use cases within the org. So engineering, customer success, upbound marketing, use cases. I think are mainly where we're seeing adoption right now in the company.

Michael Finocchiaro

Okay. But it has okay, at least I answer first and then I've got to follow up.

Elissa

Yeah, no, I would say a metaphor we mirror that fairly closely. think one of the most useful things for us has been more on the productivity side, especially when it comes to creating just like documents, and so on. ⁓ I think though on an engineering side, not to get into how we do that, but it definitely leveraged more sparingly there because of the nature of the the work we are actually doing on geometry is not as effective in that area.

Michael Finocchiaro

because it's not text-based, it's geometry-based. But are you seeing that... Well, so you guys aren't... Okay, because I was wondering if how AI might be changing the model. Up to now, we've been developing with the V or we've been doing agile, right? And it seems to me talking to some other startups that those models are actually sort of merging in a way because the interactions between developers are so much faster.

Elissa

Yeah, yeah, exactly.

Michael Finocchiaro

developers with an agent next to them is sort of going faster than even maybe a team of developers because you can pop up in UI, then you get a bit. In other words, it's sort of an 80 20 thing that you can get 8 % of the way to something that's working. And that other 20 % will take you 80 % of the time to get right all the corner cases and, stuff like that. And it seems that that's changing a bit or that could have a big impact on agile and, and, waterfall and, or, right. I just wonder what are your thoughts on that? Elissa or Rui.

Rui Aguiar

So yeah, I think what happens is you get a team of editors instead of a team of writers, which is kind of interesting where, you know, beforehand you have your coders and you have your development team and they're constantly just like writing, turning out code, running these agile methods. And then, you you get something like log code or these multi-agent system spin-up or, you know, some of these agents like software engineering. And what happens is the engineer ends up becoming like an orchestrator and like moving to a higher level. sort of complexity where they're like managing these agents like working through these kind of like high level problems. ⁓ And then basically you, you know, I guess that's what you mean by. OK, then this becomes a team because it's one person and they're coordinating five agents. And one's working on the front end, one's working on the back end, one's doing database migration, one's doing your scrum meeting. I don't know if that actually changes the model. From what we've seen, Agile still works OK. It's just a question of, OK, now you get way fewer people. You're coordinating higher leverage work and your cycles are faster. But you still have your ticketing boards. You still have some of your scrum meetings. It's just a faster iteration cycle. You're getting 2x, 3x the amount of work done in the same. The processes are still okay though because like you have external stakeholders like customers you got to work with, project management, and there does still need to be a way to like collaborate between these cross-functional teams.

Michael Finocchiaro

How about your experience too?

Elissa

Yeah, I agree with this. No, I agree. I don't think it's like fundamentally changed the way ⁓ engineering and code is being written. it, not in the way you're describing like the agile versus waterfall. It hasn't meaningfully changed that, at least on our side.

Michael Finocchiaro

Okay, so ⁓ in the Cosmon and the Metafold, how does the user encounter AI when they're using the tool? Is it part of the UI or the philosophy part of maybe the trained models underneath the covers? mean, how is AI integrated into the product as is?

Elissa

Are you, you wanna start?

Rui Aguiar

Yeah, I can talk about this. we're AI native, know, we're, basically the entire thing is AI like end to ends for whatever that means. But really it's a really like, it's, it's, it's immediate. You know, a lot of our, we're very similar to like a cloud code for, you know, engineer. This is something where we kind of view ourselves as the AI bar for hardware engineering or engineering use cases. So generally, you're interacting with it. Whenever you want to use the product, it's like an AI native interaction. Obviously, there are other things here that are important, like you're running a human in the loop process. There are different processes in place where you're interacting with a design software, a simulation software, and then you're interacting back into Cosmon. Generally, what happens with us is we try to keep things, I believe that all great products are simple. We try to keep things as simple as possible. We put up something that we try to hope can work for engineering processes. can get a lot of ways we do that training on engineering data processes, picking the right use cases that AI can work with. You can interface with it. We try to get at least 80 % of the way on a lot of tasks that aren't like incredibly difficult for engineers, ⁓ automating away a lot of that work, getting a lot of that work done more easily for the engineer and the engineering team, and then speeding up our kind of design to production life cycle. to your question, the whole thing is kind of, know, AI is an AI-native company. The entire thing is AI. have bunch of deep learning researchers on staff. Like, so when you're interacting with the product, you're interacting with an agent, you know, ideally, basically as close to 100 % of the time as possible when you're.

Michael Finocchiaro

Interesting. I suppose it's a bit different at metaphor.

Elissa

Yeah. So this, this is where I think Rui and I completely diverged, uh, that our product is not an AI product at all. Right. So our product is really a, uh, a geometry analysis product, but where it has this intersection is that it, and it kind of comes to the title of this, um, this episode, I suppose is about infrastructure. So we are one of the challenges of 3d geometry is that the representation of that 3d geometry comes in all different forms.

Michael Finocchiaro

Yeah. Mm-hmm.

Elissa

And if you want to do meaningful learning on this, you really need a unified way of interacting with the geometry. So this is something we have developed a way of both representing geometry and converting it into effectively a feature vector that you can use as the input to a machine learning model. So ⁓ for us, if you're interacting with AI in the metaphor universe, you're probably

Michael Finocchiaro

Hmm.

Elissa

⁓ you're probably running like a fairly vanilla machine learning model on top of our geometry representation. So maybe you have a large corpus of 3D models, you're trying to understand, you know, groups of similar geometries within that data set. This is something you can get by running them through Metafold, getting the feature vector, and then doing some learning on top of that. So really ⁓ AI and machine learning is applied on the outputs of Metafold and we become the kind of infrastructure for converting 3D geometry into something you can learn from.

Michael Finocchiaro

And when you said the different formats, you're about parametric versus implicit, which is nerves versus all these other, those kinds of things.

Elissa

Exactly. Yeah. Well, and critically, most of the time with 3D geometry, there's some discretization that happens when you simply represent 3D geometry on a computer. So you're probably talking about a whole bunch of triangles that represent the surface of the model. then exactly. then you may end up, if you're trying to just run some kind of learning algorithm there, you may end up learning more about the representation than you do about the intrinsic shape itself.

Michael Finocchiaro

Mm-hmm. rasterization, right? the actual geometry.

Elissa

And this is critical. It should not matter how you represent the shape. What matters is the shape itself. And this is with engineering, the true challenge in engineering is that the details matter so much. Sharp edges, features, holes, it all matters. And it all has this kind of latent expertise embedded into it that is very difficult to learn.

Michael Finocchiaro

It makes me want to ask, ⁓ so you guys both have to interact with the legacy systems, the systems that have been there forever, the Zemus PTC DS universes. So how do you guys fit into those workflows so that the user doesn't have to interrupt his work and so how does that actually work so that it's ⁓ inside the workflow of the engineer, Elissa?

Elissa

Yeah, I ⁓ I think you're right. We have to interface with existing tools. And so this is why our, ⁓ like our offering is fundamentally API based. It just means it's extremely flexible. We can be inside existing tools. We can create applications and here's like, here's where AI is amazing. You can spin up a web app very, very easily, things like this. So ⁓ I think for us, it's all about flexibility, you know, treating different file formats, different representations, but bringing them into. the metaphor representation and ⁓ leveraging that to get the information that people are looking for, but then also being embedded into existing tool chains and workflows so that we're not like adding a thing and adding a step.

Michael Finocchiaro

Then when you're done, does it get pushed back into a system of record, whatever you did with Metaphone?

Elissa

Yeah, most likely. mean, it's very case by case, but yes, exactly. Like we try to come back into the format that people are expecting.

Michael Finocchiaro

Whereas Rui, you're more acting on top of between the systems of engagement and the systems of record, right? Because you're working with the CAD guys and you're also having to directly interface with the system of record,

Rui Aguiar

Yeah. Yeah, so we kind of model ourselves off of a lot of the sort of more common agentic tools that are workflow orchestrators. So again, it's not necessarily like you're bolting onto something in a very similar to probably Elissa here, it's API based, you you're interfacing in and out, you're working with these tools, you're exporting things, you're importing things in place. With us, like, yeah, we're like workflow orchestration. So we have to connect to the tools because that's like how you do engineering, right? Like you're not going do engineering with your Creo or your SolidWorks or your Ansys or something like that. Like that's just not going to happen. So we kind of have to do that. The challenge for us becomes, you know, maybe similarly, like, how do you think about interfaces and boundaries here? Because, you know, this is, it's not text-based, there's metadata, there's information out of these systems. Like, for example, we have a customer who we're talking to or who we're working with and they're like, well, I want to use this across.

Michael Finocchiaro

Ahem.

Rui Aguiar

specifically Ansys and SolidWorks, right? But they're like, what can I do? Like, how can I use an agent to take the output of a simulation and then use that to go modify geometry or think about how I can perform geometry modification upstream or downstream of the results and perform these changes? ⁓ And so when you think about integrating into existing tooling, it's not just plugging into one tool or plugging into another tool. It's like, how can you take the context out of both of these types of tools, understand like how one can kind of implicitly reference the other, maybe with help from the engineer of like, hey, here's what I to look at, here's kind of a judgment call or perspective in terms of how to think about this, and then kind of work across both of these. And then when you think about an agent, when you think about development, it then gets really hairy of like, well, how do I even think about the boundaries between these two? When do I know how to switch? How do you build that into a product line? And so ultimately, think it becomes like, number one, we want to be basically from an interface perspective, you want to be as seamless as possible on top of an engineering workflow as an orchestration tool. Number two, you have to coordinate multiple forms of information, and that's still an open question as to how to actually perform this, which we're trying to solve. And then number three, you have a user experience question here. Even if you're working on top of these, even if you're bolting on top of these, how do you move across tools? How do you move across orchestration in a reasonable way that is not super error-prone or that an engineer is amenable to? Maybe it's human in the loop. Maybe it's some other things. But that's we think about running on top of it.

Michael Finocchiaro

Interesting. ⁓ I usually throw a question about this time about for the next generation of engineers, what kinds of things they'd be focusing on. But also makes me want to think about how engineering itself is going to be evolving more towards prompting than engineering, but with an understanding of the physics, right? So it's kind of interesting the way that the, I mean, if you, it's almost impossible to even look. three or four years down the road because the things are just changing so fast. But it seems to me that, well, what do you guys think are the critical skills that the kids that are in school and looking at the job marketing saying, my God, might as well go to theater because it'll be just as easy to get a job from an AI in theater than it will be in engineering, right? ⁓ What kind of advice would you give to them, Elissa?

Elissa

man, that's a such a tricky one. And I'm like, I'm reluctant to go on record, giving any advice here. ⁓ Yeah, exactly. But I mean, like something actually, what you just said, Rui was interesting, like you have these agents, and they, they can make, make, ⁓ they can kind of not automate, they can do a lot of the work of the engineers that, you know, previously, people went to school to do this. But you also said something, which is that, like, at some point in that flow, like you need to

Rui Aguiar

We'll have to scrub it after.

Elissa

out and the engineer needs to like bring their expertise and come back in and direct, right? And so I, maybe, I don't know, like, I still think that a kind of classical training and an understanding, a true understanding of the physics and the engineering principles that are at work is so important because this is not, this is not like these agents, like they don't understand the 3D world that we are sitting in right now. They don't have the understanding of a child even about the 3D world. So we can't outsource our engineering to them. Like we need to maintain ⁓ the integrity of the field. I'm really curious to hear what you're going to say, Roy.

Michael Finocchiaro

Hahaha

Rui Aguiar

Yeah. So look, I mean, I think like it's actually, I generally agree with this. would say like even in software, like I wouldn't hire someone who just missed how to use an agent. Like you need to first principles understand like why you're coding something or building something out. I don't really think that ever goes away. This is probably especially true in hardware where you have like way more expensive costs down the line than like if you ship a broken website or something. And even for software. So, I mean, like, you know, I've, I've had interviews where people come out of college and talk to me and they're like, Yeah, I know everything. Like I just have an LLM feed it to me and that's how I did my job. I'm like, okay, this is not, I don't think this is going to work out. I'm sorry. Maybe you can go somewhere else with respect to knowing everything, but I think similarly, like, you know, it's again, it's a spectrum. I think that it is going to be important to be able to learn and understand how to use agents for different types of workflows. Not everything in engineering is 3D geometry, just like not everything in software is coding.

Michael Finocchiaro

Ha Ha ha ha.

Rui Aguiar

And I think agents are going to be useful for different things across these PRM specifications, readouts, maybe some basic analysis, interpretation, these types of things. ⁓ Even modifications, but regardless, it doesn't really matter because at the end of the day, like if you can't ground the thing and like a first principles understanding of physics and a first principles understanding of engineering, it's way too risky to, to, to, you know, ship a product or build a product out doing that because you're just way too prone to make mistakes. So I'd actually maybe in a disappointing way, cause it's less controversial. I'd agree a lot with what Elissa is saying here. ⁓ sorry for the podcast listeners, but, ⁓ I would say that like first principles, understanding is still incredibly important. It is still incredibly important in stuff like software. You need to understand this. need to be formally trained in really thinking through logically from a first principles perspective. Even if you're using an agent, is what this agent's saying actually right? These things are stochastic machines. They are pretty good a lot of the time at a lot of things, but they are not 100%. And they will never be 100%. And so you're always going to need that grounding. think human in the loop is always going to be super important. I would say that like, from an educational perspective, keep that in mind, keep that human in the loop is going to be really important and you're going to need to understand all this stuff and engineering is not going to go away. But also try to educate yourself with newer technologies that are coming out, understand that this field is evolving incredibly rapidly, understand that, look, what's possible today? There may be things in five years that are possible now as you're finishing a degree up, if you're starting college or something like that now that you didn't think were possible when starting out. So I would say keep it as a hybrid, flexible standard. Don't neglect the kind of class principles here, but also understand how to use AI, understand where it might be useful, understand where the limitations are, try to incorporate it where you can in a workflow, and try to stay on top of the developments in the field.

Michael Finocchiaro

Do you guys feel that with the fact that these tools are getting somewhat easier to use and making it more accessible to more people, that the engineer of the future will be, forcibly, much more multidiscipline than he or she is today? Because today you can just live your whole life as a mechanical engineer just doing design. But I think tomorrow you're going to have to know a little simulation. You have to know a little bit of manufacturing. have to know... what a milling machine is and whether that's actually you can get the miller into it or you're gonna have to make a decision that no, that's gonna be additive and then I'm gonna use generative design to do that. Don't you guys agree with me that there's gonna be some modifications to the we're engineering where you're not gonna be able to just say I'm really good at this one thing that you're gonna force, you're gonna probably have to say I'm really good at these three or four things or maybe I'm way off.

Rui Aguiar

I would say like, you you see a similar dynamic to this. Again, I keep having to do this, but software is like the one field that's actually been disrupted significantly by agents. So I'm kind of extrapolating into like, okay, this is probably what's going to happen to these other fields over the next decade or whatever. ⁓ I would say like, you see kind of two forces at play here. You see like upskilling and you see democratization when you like interface with an AI. Like the first thing that AI and agents and stuff really good at is, you know, number one, ⁓ You can democratize access to a lot of things here. So there are these other, there's this company called flux. They do like AI for PCB design use cases. And like, what's really interesting about them is their baseline customer is actually not an electrical engineer or like a designer. It's someone who's like a mechanical engineer who like wants to break into these things, but like doesn't really know how to kind of develop or can't get access to the data and wants to just spin stuff up and test it. So I think to some extent, like there may be like a higher, like there may be sort of a rising tide across all things here where it's like, Hey, now can give you a baseline understanding of simulation, of how to think about design, a PLM, or something like that, right? And that'll just naturally come with an engineer. That may not even be something you're forced to do. It's something you just get if you use these tools well enough, and they can integrate well, and so on. ⁓ And then the second point is, I do think probably the... And this is an interesting question. The question of what is the baseline that matters ⁓ become interesting? Because I think in software, what you see is... You know, the, new grad job market or the sort of like less experienced job market gets contracted significantly because you can get a lot more done with like a more experienced person who kind of knows this stuff. Now, you know, hardware engineering and mechanical engineering is obviously kind of a lot broader in some ways than just writing code. You have to run through a bunch of different systems. there may still be some use in specialization, but I do wonder if like, you know, someone who is more experienced to us use these things a bit and they can get up-leveled immediately with like an agentic tool or something like that. could be more productive and could cause, you know, hey, we can just get one person to do a lot here. Again, because there are much more limitations in engineering and software with AI than there are with, you know, something like coding, where it's one-shotting the thing end to end, that probably doesn't happen in the dramatic of a scale. But maybe you see some kind of like compression forces like that in action. We'll have to see.

Michael Finocchiaro

Elissa, you think, when do think we'll see the open AI moment for engineering then? When will geometry be able to be, have its own model and there'll be like the open AI moment. Do you think it'll be in the next couple of years or you think, you know, this is decades ahead?

Elissa

you Yeah, no, I wouldn't make that kind of prediction. I agree with what Rui has said. fact, obviously not all engineering is 3D, but 3D is a big part of engineering. And ⁓ there's huge challenges here. We simply do not have the kind of data that you would require to have an open AI moment for 3D geometry, in my view. So ⁓ yeah, not going to make that prediction here.

Michael Finocchiaro

Well, then let's just think about the fact that we've gone from 2022, 2023 with the advent of OpenAI. We went through the MCP period and then now we've hit whatever the hell that multbook open cloth thing was. Are you guys as bullish or skeptical as you were in 2022, 2023? And where do think it's going? Any ideas on where we end up? There's two of you now. for some reason. ⁓

Rui Aguiar

Yeah, my computer just crashed. I don't know why. Sorry about that.

Michael Finocchiaro

⁓ So

Elissa

Yeah.

Michael Finocchiaro

like, do what do you guys think is coming next?

Elissa

Yeah, I mean, I'll say something like super obvious, which is that it's very difficult to tell what is going on. So there's huge amounts of capital flowing into this industry that are, you know, absolutely affecting public perception of what's going on. And then also like expert perception of what's going on. So I think it's very difficult to actually understand. And maybe something we'll come to is like the actual adoption and the actual digital maturity of organizations who might be the people using these tools. Cause I still think that that's a ways off. yeah, hard to tell what's going on.

Michael Finocchiaro

So you're still a bit skeptical basically.

Elissa

I, well, yes, I am like, I am very skeptical. mean, I am not skeptical that AI is a powerful tool and incredibly disruptive. This is, I don't think, ⁓ under question, but yeah, no, I think this is very true, but I think like where it is going to be extremely disruptive is, is still, ⁓ is still unclear. And I don't see massive disruption coming.

Michael Finocchiaro

a sane person could say otherwise.

Elissa

in the engineering or manufacturing, like we're in manufacturing. in manufacturing, like people are producing physical things. are, ⁓ there's only so much that can be disrupted there in my view.

Michael Finocchiaro

Are you a skeptical Rui or a bit more optimistic?

Rui Aguiar

I adoption curves, ⁓ also is there an echo? I don't know if I'm in this twice, no, I hope not, but okay, good. Yeah, my browser crashed, sorry about that. ⁓ look, think the adoption curves will definitely be slower on a lot of things here. I think will be, know, like organizations will not adopt for...

Michael Finocchiaro

No, you're good, you're good.

Rui Aguiar

Even if the technology is capable, which obviously there are limitations for engineering in the first place, like organizations will not adopt for years afterwards for a variety of reasons. Data security, IP readiness, skepticism. ⁓ you know, process training and so on. And I think this is probably be one of the slower moving fields to adoption. This, yeah, as Elissa said here, I manufacturing is especially hard. the less you move at a conceptual where like you're going and the more you move into like physical manufacturing, that is really difficult to kind of displace. And I would probably agree with like, look, that this is going to take a while. And even for something like what we're doing, it is likely to take at least years of adoption into these larger organizations, even if the technology is working in the first place for some of these types of workflows. ⁓ I would say like, That said, there are large pushes by these companies to adopt AI in different forms, in software, in customer success, sometimes in engineering, sometimes in coding, sometimes in these other use cases. And what we'll probably know in a few years anyway, from a ShakeOut perspective of how this is going to work, I predict that there will be a large push for this. It will be successful in some places. It will be less successful in other places. The market will kind of naturally figure out, okay, here's specific use cases where AI is good for. Here's where it's not. I do think this is a question of at least a few years, not like a few months or next year, even in two years. But I do think that like generally like markets are not that inefficient. Like there's a lot of capital going into AIs as we've said, like there's a lot of experiments being run. Some of these will shake out. Some of these things will work. It's really about the ability of. the company to figure out where there's true demand, what's at the edge of the feature complete set of things that are useful to an engineer that are also technically possible, delivering that to an organization, and then seeing tangible business ROI from there. It's probably where you'll see adoption. I think everything else may struggle.

Michael Finocchiaro

⁓ So the other thing I want to talk about is ⁓ digital transformation in general. And when I look at digital maturity of companies, I think of like the least mature ones still using email and ⁓ as collaboration tool in Excel for almost everything else. ⁓ And I'm thinking at the high end, like I would give that a ⁓ one and a five would be fully agentic autonomous digital twins. And I don't think anybody's there. And I think it'll be a while before we're there. ⁓ So first question would be like the customers or you guys are already are using Metafold and they're already using Cosmon. Do you think are they at between one and two or do they have to have a certain maturity level to even think of using your tools? Like if they were only on one, they'd be like, I don't even know. I don't can't even conceive of using your tool. I mean, just where do you guys see your customers on that ⁓ continuum?

Elissa

Yep. You go. You go.

Rui Aguiar

say that... Okay, yeah, I'll go first. ⁓ Yeah, it's a spectrum. I would say that if someone isn't even bothering adopting cloud and has no use cases on top of this and is 15 years behind, it's very difficult to leapfrog that. Even if the technology is sufficiently disruptive, it's like, this is not an early adopter and some people only use it, which is fine. And you get some people on that. I would say that...

Michael Finocchiaro

Yeah.

Rui Aguiar

You know, on the other end of this, what's up? Yeah. Like a, like a two or three. would say that like, generally you need to have some degree of comfortability with like automation you needed for the, for us at least, you need have some degree of comfortability with cloud. ⁓ and those would be the two main ones. And that's probably like, yeah, about a three out of five or so from an adoption perspective, anything that's like less trustworthy than that or is, you know, has IP risks or something. I mean, IP is important, but.

Michael Finocchiaro

So you're considering like two or three?

Rui Aguiar

⁓ is like, okay, we can't use cloud, can't exfiltrate anything, we can't touch behind a firewall. It's just challenging to adopt something like an AI with because unless you're deploying something on premise or just unless you're deploying something deeply within an embedded ⁓ infrastructure network. ends up just being a challenge to spin out to. So I would say three or above is good. If they're hosting their own LLMs, which many firms are now, that's great for agent deployment. At least some degree of comfortability on cloud, at least some degree of comfortability on automation is probably what we look for for the sweet spot of ideal customer adoption.

Michael Finocchiaro

Have a great day.

Elissa

Nice. Yeah, it's similar. know, we, ⁓ because we work in manufacturing, we, we see a lot, ⁓ like a big range, I would say. And we've actually been, we came out very bullish on cloud and we've had to adjust somewhat. And so in some cases we can be completely air gapped ⁓ from the internet. ⁓ Yeah, you we do a lot of work in footwear. Footwear is interesting because it's a very traditional industry. And so a lot of the time, people are just like making physical things and testing them. And so this is a very early stage of digital adoption. you know, we can show them that actually they can do so much digitally. And ⁓ it just kind of is mind blowing. Typically, that's not the great fit, though. Like, it's better when people are further along a digital adoption. But it's also exciting. It's honestly exciting to be helping some of these companies, you know, digitize and move from two dimensions to three dimensions where you get so much more information. Like this is also a big unlock for a lot of people. I have not encountered any companies in the manufacturing space that are past like a four, even a three maybe. It's more traditional.

Michael Finocchiaro

Well, my thesis is that using software like Metafold or Cosmon kind of moves the bar though. If a customer does take the risk and say, okay, I'm going to actually use these tools, is there sort of an epiphany? Have you seen an epiphany where the customer says, my God, if I didn't have all these data silos and my customers, my engineers actually talk to each other, they didn't hate each other and different departments were communicating and I didn't have data siloed in each department, they couldn't talk to. Are you seeing the epiphanies where they're saying, wow, if we fix that, we'd be a better company and we'd actually make better products. And thank you, Metafold, and thank you, Cosmon, for bringing that epiphany to us.

Elissa

Yes. Yeah, I love this because we have exactly seen this where people kind of begin to understand this opportunity and then start thinking bigger. Now the challenge can be that you might have someone who lashes onto this idea and like sees the vision, they see the future, they see what the possible could be. It can be so difficult to get them, get the buy-in from the rest of the organization depending on the scale and the structure. ⁓

Michael Finocchiaro

That's awesome.

Elissa

It makes me really think about change management and how difficult that can be on an organizational scale when you're dealing with these large global enterprises. It's amazing. Especially coming from a startup where you just like, move quickly, like so quickly, right? And then you enter these large organizations that just are so slow.

Michael Finocchiaro

Mm. That's awesome. Rui, have you had that same experience?

Rui Aguiar

Yeah, I would say similar. You usually this happens on the PLM side for us mostly when you start thinking about information, architecture, orchestration, and so on. ⁓ I would say that this is something we've seen before, like, hey, I want to go connect, especially because we're workflow orchestration. It's like, well, can I connect this to this and then pull this thing out? And then like, what if we could have a better search over this version of our PLM or something like that is kind of where people start to put the pieces together on. It's interesting. You know, I would say that like the big things are Number one, yeah, the cross-team coordination and looping in other departments is probably sometimes where that can get difficult. Within a team, within an organization, within certain parts of a PLM, this can be useful, but working across an entire global enterprise is slow. It takes a long time. There are the light bulb moments, but I still think even if there is...

Michael Finocchiaro

you

Rui Aguiar

internal conviction, there's still bureaucracy, there's still sort of pushing through on things, which becomes a challenge. That said, you know, it's pretty clear this is where things are going. think like better search, better information retrieval, better use cases of data and coordination of data between teams and the ability to think about how to move that data between teams is the future of where engineering data management is going. How long is it to take to get there? I don't know, but like pretty clear that like people are seeing this in a lot of fronts. Even I think, you know, across the list and we have two pretty different products. like people can start putting the light bulb on, like, oh, hey, wait, what if I could digitize more? What if I could have a sort of better digital thread or so? And think people kind of get turned on to that a bit.

Michael Finocchiaro

Yeah, what if I could, you know, you have to. We actually did have a question from Lisa Finwick in the chat and she asks, how is a company's IP protected when using these two tools, Cosmon and Metafold? What data sources are being used for the AI to operate on?

Rui Aguiar

I can talk about this first. Yeah, so I mean, we run both cloud and air gap deployments depending on data sensitivity levels. you can point basically, like we have like a, it's an agent, but like the agent can get deployed wherever you just kind of need something to point it out as like a brain or like, you know, to drive it. We have our own models that like we can develop out and can get bolted on onto a network. So you don't have to leave if you don't want to, you know, some people are okay with cloud. Some people aren't okay with cloud. We try to be flexible to both. ⁓ The second thing is ⁓ we don't upload or export the models or anything like that. We ⁓ don't touch a lot of things. We touch a little bit of the metadata, really. ⁓ We get audited. We get tested, stuff like that. But generally, what I would say is we don't export the models. We don't touch a lot of the IP outside of a lot of local processing on computer for these use cases. And the second thing is if you're super sensitive, like if you're like, hey, cloud isn't OK, ⁓ it doesn't really matter where the AI gets deployed. A cloud is just a computer. So if you have something in your own network, if you have an intranet, if you have the level of maturity where you can deploy that, you don't have to leak data at all. You can just run everything behind a firewall. It's just about deploying it in the right network or right local computer set.

Michael Finocchiaro

Very cool. What about you, Lisa?

Elissa

Yeah, we obviously take IP extremely seriously and for manufacturers, their products are their IP. So this is very sensitive data. ⁓ Because we don't have one model that we are operating on, there's never a sense that we're taking people's data and using it to improve or train a model. People can train their own models on their own data. And in fact, this is a huge source of potential competitive advantage for manufacturers who have a large corpus of historical product data. So we're always really excited to help people do this, but it's always just their data that they're operating on. And we're very careful to keep that, you know, that lives with them, not on our servers.

Michael Finocchiaro

Understood. ⁓ Actually, I'm kind of done with the questions already. That was kind of quick. ⁓ Yeah, no, it's great. ⁓ Any parting thoughts ⁓ from the two of you about AI, about your products, about where the market's going or where engineering's going?

Elissa

Amazing.

Rui Aguiar

All right. I think it's too fast. I think you just got to stay on top of it. I mean, as of today versus a year ago, it's a completely different world or two years, forget two years ago, right? I would say that like parting thoughts on this are just, regardless of what you're doing, what you're working on, it's disruptive technology. We don't know where it's going. Nobody knows where it's going. think people can guess, but staying on top of it, making sure that it's front of mind is really important. And I think the best way to think about how to really deliver value, even if you're not sure right now is just understand where it's getting, you know, what's working and then thinking about that for your own use cases.

Michael Finocchiaro

⁓ huh.

Rui Aguiar

and the analogies there is useful.

Elissa

Nice. Yeah, agreed. And it's a, it's a, it's an interesting time, an exciting time. Um, hard to keep up, but like, yeah, I think there's so much opportunity and potential here that it's an interesting time to be building a company. And I think an interesting time to be adopting new technology as well. Um, you know, our approach at Metafold is always to try to help people scale and kind of liberate and, and amplify their engineering expertise, not to, not to replace it. ⁓ And so I think that's, yeah, that was the message I'd want to leave with on Metafold.

Michael Finocchiaro

And if I may, one more, another question. It seems to, it's a very competitive landscape, right? Cause there, I found on my database, I have 500 startups in engineering software and international AI. It's astounding how many of they are. And so it seems that you guys have both chosen a strategy of kind of picking one industry. know, Rui, you're doing hardware and Alisa, you said you were focusing primarily on footwear. Is that? the survival strategy or is that just because it most interested you? Is that the way that you're able to carve a competitive mode because you're able to solve the problems very specifically to those industries? I'm just saying in terms of the strategy of how for other people that might be listening and thinking about, hey, I've got this great idea, but what do do with it?

Elissa

Yeah, for us, it's been important to pick a sharp focus. don't just work in footwear. have a few different industries we support, like electronics as well and aerospace. But ⁓ I think it's important to be very specialized and very sharp. that's the, it's a bit of a survival strategy in the sense of like, how do you begin? How do you enter a market and begin? And that's not really like a technical challenge. That's a business and go to market challenge.

Michael Finocchiaro

Interesting.

Rui Aguiar

Yeah, I think it's a forcing function. mean, you got 10 person teams, something like that. Like you're not going to be able to take over the entire world of engineering. I mean, as much as you may like to. ⁓ It's really just like, look, there's only some things you can do. You're really only going to get business if you do those things incredibly well or better than.

Elissa

Yeah.

Rui Aguiar

alternatives. And so like you have to narrow your scope to be able to focus on like, all right, we need to do like maybe two or three things max, like super well. That's it, right? You know, and do that for years. And then if you get lucky and you get enough, can kind of expand and specialize here. But definitely for a startup focus is incredibly important. I wouldn't recommend doing more than like two things probably at a time, even that's probably pushing it. Some CEOs are like, look, just do one thing and do it extremely well. But that's kind of where we sort of balance here.

Michael Finocchiaro

Hehehehehe That's awesome. it's only, we're almost in March, but where can we see you guys in public? Where are you guys gonna attend some conferences like the Threaded Conference in Miami, for example? Where can people come and meet you guys and see your product live?

Rui Aguiar

You Yeah, so you can see, not me, but my co-founder will probably be at Threaded. I'll be at GTC. ⁓ So I'll be at GTC. I'll be out there. I'll be in a couple of other conferences across Silicon Valley, Simulation World, Ozen, GTC, NAPHEMS, and then also Threaded should be out there, and then a few in April as well. So lot of the engineering events conferences, you can probably find a rep from us there. Come say hi.

Michael Finocchiaro

No, a bit cool.

Elissa

Nice. Yeah. Yeah, we tend to attend more manufacturing focused events. So we're doing one in Portland a couple of weeks. Yeah, we will be there. That's a little later, but more sooner we'll be at in Portland. ⁓ And then also at Hanover Messe in April. Yeah. Yeah, yeah, yeah. I'm excited for that. Yeah.

Michael Finocchiaro

How about you, Alisa? ITMS. ⁓ over here in Germany, nice. Very cool. Are you going to go to CDFAM in Barca too?

Elissa

Not planning to attend this year. do love the CDFAM events. I've given plenty of talks with them, but not on the radar for this year.

Michael Finocchiaro

So I'm thinking of going to the Barcelona one. haven't met Duann yet, so I'm thinking, yeah, it's going to be a lot of fun. Well, that's awesome. yeah, I think it was great conversation. I appreciate you guys both taking time out of your busy schedules. I hope I get to meet you guys in person at some point during the year. So anyway, thank you very much. Thanks to the audience for tuning in. And I'll see you in the next AI Across the Polyglycical podcast. Thank you, everybody.

Elissa

you should go. Yes, they're great. They're great.

Rui Aguiar

Travis.