Stop Running CAE Like It's 1995 — with Key Ward and EMMI AI

Episode Summary

The episode delves into the evolution of computational analysis and engineering simulation through the lens of EMMI AI and Key Ward, two innovative startups at the forefront of leveraging artificial intelligence (AI) in product lifecycle management (PLM). Dennis Just from EMMI AI shares his background as a former CEO of smallPDF and his journey to founding EMMI AI with Johannes, focusing on applying deep learning methods to large-scale engineering simulations. Asparuh Stoyanov of Key Ward discusses the company's mission to break data silos in engineering-heavy industries by unifying scattered data from various solvers and sensors through their platform, Key Ward Hub.

The episode highlights two critical insights: first, the distinction between physics-based models and surrogate models—EMMI AI focuses on fundamental AI-driven methods for complex simulations, while Key Ward emphasizes workflow automation to streamline data management. Second, both companies underscore the speed and efficiency of their solutions compared to traditional approaches, with Asparuh noting that they can deliver value within days or weeks, whereas conventional physics-based models might require years.

For PLM and engineering professionals, the key takeaway is the potential for AI-driven simulation tools to significantly accelerate product development cycles and improve decision-making processes. By integrating these advanced technologies, companies can achieve more efficient data utilization and faster innovation, ultimately enhancing their competitive edge in a rapidly evolving manufacturing landscape.

---

Full Transcript

Michael Finocchiaro

And we're live. This is Michael Finocchiaro on the AI Across the Product Lifecycle podcast. I'm joined today by Dennis Just of EMMI AI and Asprey Stoyanov, I hope I got that right, of Key Ward, ⁓ two really exciting ⁓ startups in the simulation space. ⁓ So Dennis, give us a little bit of your background and about EMMI AI

Dennis Just

you Thank

Asparuh Stoyanov

Correct.

Dennis Just

Yeah, more than happy to and thanks for having us Michael. Personally, studied economics and mechanical engineering and have been professional gamer before and have been on the I'd say, entrepreneurial train for the last 15 years. So, build scaled exited companies before founding me was CEO of small PDF, one of the largest companies, but for companies in Europe.

Michael Finocchiaro

My pleasure.

Dennis Just

in the document management value chain as an example. And then two years ago, I had the luck to meet Johannes, which is the brain behind everything we're doing. And Professor at the JQU led the large scale simulation group at Microsoft Research and was responsible for the Aurora weather model to apply all the learnings that the guys had in weather modeling and ultimately moving from numerical methods towards data-driven AI models in weather modeling towards engineering. And this is where we started with Emmys. very much have a fundamental approach towards the problem. The deep learning methods frameworks and in very large scale models in happy to talk more about it, Michael, in the podcast.

Michael Finocchiaro

Fantastic. Yeah, thank you. By the way, I am looking at comments. you guys listening in the audience wish to make ⁓ questions, either of them, please do so. Asprey, tell us a little bit about you and about keyword.

Asparuh Stoyanov

Sure. Yeah, first, it's a pleasure to be part of the podcast. Yeah, the topics are very interesting for me as well. ⁓ My background is mechanical engineering. ⁓ As most of the people in our company, we come from mechanical engineering, engineering background. ⁓ And this is for half of my career. And essentially around 10 years ago, ⁓ I switched for focus into data science, computer vision. And essentially that's how we started keyword. We were coming from the aerodynamics background in automotive field. And yeah, we saw this potential to use all this data to train surrogate models that can do these evaluations much faster. However, during the process of building those surrogate models, we discovered another problem in the industry. that actually many companies, they do have the data, but it's everywhere ⁓ and it's siloed, it comes in many different shapes. And many companies have difficulties making something useful out of their data. And that's how we actually started our platform, Key Ward Hub, which is tailored for engineering and data heavy industries. ⁓ And with the main goal to break data silos, unify scattered data coming from variety of solvers. or sensors both from R &D testing production and essentially replacing ad hoc manual workflows with Excel, Python ⁓ and yeah all these more isolated tools.

Michael Finocchiaro

Some of the people listening might not know what the difference is between like a physics-based model and a surrogate model. Maybe you can just give a half a second explanation. I think you guys are, Dennis, you're more working on the fundamentals, right? Whereas Asparuh is doing the surrogate stuff, is that correct?

Asparuh Stoyanov

More on the workflow and automation for us. ⁓

Dennis Just

I mean, it's, so I think it's.

Michael Finocchiaro

⁓ go ahead, ask Brian

Dennis Just

Yeah. And I think so there's a natural,

Michael Finocchiaro

there Dennis.

Dennis Just

there's a natural, I think evolution, right? Whereas I think in my head, surrogates are always models that solve one problem for one company and, let them optimize whatever they want to optimize. I think the path towards foundation models in engineering is I think what we're trying to pave them with tech that we have. So go further and train models that kind of cover a whole engineering.

Michael Finocchiaro

Asprey, you had a definition for us, I think.

Asparuh Stoyanov

Yeah, can you repeat the question please, Tobi?

Michael Finocchiaro

What's the difference between physics-based models and surrogate models?

Asparuh Stoyanov

⁓ Well, fundamentally surrogate bottles are kind of a, the name says it, a replacement for the physics-based bottles. ⁓ Trend on data can produce much faster results. Now we have this buzzword of physics-informed ⁓ neural networks, which essentially tries to both data-driven and physics-based into the same thing.

Michael Finocchiaro

Okay. No, it's just because I think that there's, mean, AI is moving really fast, but I think the simulation space is moving super fast too. So some people in the AI's might not be caught up. I just, put the, in the comments of the post, I put a link to my, what I published yesterday, I had this call, which was my top five trends in simulation. And the first one, ⁓ I think it's more what EMMI AI is trying to solve, right? The physics-informed neural networks replacing the traditional solvers. Is that correct? We talked about that before the call.

Dennis Just

Yeah, so I think generally speaking, we love to go one step ahead. So physics and network means that you still have some sort of ⁓ partial differential equation within the models, which on the one hand is amazing because you can guarantee kind of physics consistency, whatever you want to guarantee from a process perspective. But when you want to train models that are extremely large scale that generalize across a variety of parameters. ⁓ Usually the physics informed approach doesn't make a lot of sense because it constrains the model with the compute that you basically have to use. And it primes the model tying towards the equation that you're trying to solve and not the generalization capabilities to enforce from a deep learning perspective. So you ultimately lose a lot of the transformer architectures give you ⁓ by constraining the architecture towards specific physics. So what we're trying to do is fully move. basically say that ⁓ transformer architectures, replace numerical solvers or physics methods. But so that you utilize what has been proven in vision and language, scaling towards very large scale models that have big breadth and depth in terms of intelligence. And that way are really capable to generalize across, for example, geometries or interparameters, because that's ultimately what you want to make repetitive revenue and to disrupt the industry.

Michael Finocchiaro

Okay, thanks, Dennis. ⁓ And Asprey, I think you're more on the AI-powered workflow, right? Because you were talking about how you're working on a... It's certain that the data flows well between the different disciplines,

Asparuh Stoyanov

Correct. Correct. So ⁓ not entirely AI powered. We do use a lot of AI agents to help automate mundane and boring tasks, such as setting up templates for analysis and all those things. However, in our platform, we put the focus mostly on the engineers ⁓ and we empower them with these AI agents to do faster. processing of their data, faster analysis to come faster to insights. But we don't provide them with a replacement for some of the tools.

Michael Finocchiaro

Okay. So if we go back in time for a minute, you remember back in 2022, 2023, when chat GPT-3 came out, LLMs became a household word and most of us had never heard of those before. You guys probably have, but I mean, think it might have vaguely heard of them. Some of us were very skeptical. Some of us were very bullish right away on the AI thing. And there is a sort of bubble effect. We're kind of waiting for that burst to happen sometime this year. But where did you guys sit when it first burst, the whole OpenAI announcement? Were you guys really enthusiastic and super bullish on it, or were you a bit skeptical?

Dennis Just

I guess I was in Europe when it happened. I mean, obviously being a deep learning machine learning community, ⁓ you write on this way for excitement because you see what capabilities it from a visionary and pioneering perspective brings to you. think, you know, reflecting to engineering, what is clear is that we haven't seen yet this GPT moment in engineering. And I think there are a lot of companies who are trying to break through this moment that then

Asparuh Stoyanov

Who wants to start?

Dennis Just

It all builds the trust and enablements to really replace classic and standard workflows and partially the incumbents or the numerical solvers to create the value that you need to increase engineering velocity.

Michael Finocchiaro

How about you, Asper?

Asparuh Stoyanov

Yeah, I agree with Dennis. So when ChatGPD came out, ⁓ I saw it ⁓ definitely as a must too. And I do believe that if we had started Key Ward now, essentially what we did for the first two years or a year of Key Ward, we could have done it probably in a week or two weeks with such powerful models that exist today. ⁓ But yeah, it definitely helped with a lot of ⁓

Michael Finocchiaro

Hmm.

Asparuh Stoyanov

mundane tasks. You know, I'm an engineer and such, sense and developer, and it's really helped me a lot with even polishing emails or writing those things that are not really my thing. ⁓ It's really excelled in that, ⁓ as Dennis said, ⁓ we're still lacking the same thing for engineering. And we believe that partly is due to the architectures of the models, but I think it's to the large part because of the data. that's out there to train these models.

Michael Finocchiaro

⁓ Well, there's that and there's the complexity of ⁓ CAD in general, right, because it's such a complex data model. But that's a nice segue because you mentioned you're a developer. The ⁓ whole vibe coding thing is ⁓ quite a trend on the web. And I've got a lot of friends of mine that are vibe coding at the moment. Or I think the name has changed now. ⁓ How has... AI changed the way you guys are both developers. So how has AI changed the way you develop? How are the teams at ME AI and at a keyword leveraging AI? you guys still in the copilots? Are you guys using cloud code? Has Agile been modified to some degree by using AI? Are you bringing agents to your scrum meeting now? ⁓ I'd just like to get an idea of how the changes that are happening there.

Dennis Just

Having agent scrum masters talk with agent product managers. Go ahead, Asparu.

Asparuh Stoyanov

I mean, yeah, I guess you want to go.

Michael Finocchiaro

That's what I'm thinking is going to happen. mean, already the PRDs are written by agents, right? And nobody sits down and writes a 20-phage PRD anymore, do they? Go ahead, Janice.

Dennis Just

No, I mean, I think for us, we have the luck and a bit too, as far as reflection of, know, maybe having started earlier, you would have kind of done it differently to be a company that is very deep tech. we develop our own architecture, model architecture, we develop our own framework. So naturally these models that do kind of more, I'd say standard software engineering, like that you refer to Michael, know, of the cloud codes of the world.

Michael Finocchiaro

That is.

Dennis Just

They do help you, but not much. mean, I'm not sure if you try to develop a very complex product on white coding platforms. ⁓ Good luck. this is the more, the more specialized you go, the less they are of help at the moment. So, mean, obviously we utilize everything that we can, but we are more trying to think and apply what's possible towards what we're trying to break through.

Michael Finocchiaro

Actually, I have.

Dennis Just

And for that we have lot of PhDs in the team that are ultimately working on fundamental modeling technologies.

Michael Finocchiaro

Okay, so it is as had an impact on the development. ⁓ How about a keyword?

Dennis Just

100%.

Asparuh Stoyanov

I agree. ⁓ It has definitely had impact on the development speed, but as Dennis said, mostly ⁓ for tasks which are kind of classic and typical software development or data processing tasks. But when it comes to ⁓ R &D and really discovering or architecting something new, that's where these models cannot really do the brain work for you, ⁓ especially if you want to innovate.

Dennis Just

Thank

Asparuh Stoyanov

But in terms of the software development part and packaging everything into a product, it definitely has helped us a lot. Yeah. And as you said, peer these as well.

Michael Finocchiaro

Well, has it changed that? Has it changed the way, well, we were talking about it yesterday on the other call about how there's some industries that absolutely have to have a V. They can't really get around the development V. And then traditional software, we've had this agile thing with the milestones and the scrums and everything. ⁓ seems like AI is gonna transform that. It can't possibly be one of those models. There's gonna be some hybridization. There's already been a lot of hybridization, but how do you guys see that? but at the same time, I think it's sort of real. I said, maybe we've got an agent that comes to the scrum meeting, or maybe there's a bunch of agents having their own scrum meeting and then giving us a summary of what they did. How do you guys see that happening?

Dennis Just

It's a tough one. mean, I think as much as I wouldn't like to fly a plane that hasn't been physically tested, I wouldn't like to use a software that has been programmed fully by AI without a human in the loop. So I think where it can help, and this is very much what we're trying to facilitate at ME, the task where human brain power is just too much and has a negative ROI on executing it. And as parallel kind of reflected right to the email, detailed description, last to have whatever you, you, you name it. Um, that is just fucking boring and it's, it's, it's not worth the money to actually do this stuff. But when it's really about being creative and doing fundamentally new, innovative stuff, um, it, it, you know, it doesn't help you. Um, because this is as you've, and it's very much relates to, think AI and engineering, right? These machines. as big as they are interpolation machines. So they will not figure out anything new that you've trained into it. It can maybe draw a relation between two data points that you haven't seen before, but it will never be capable to invent something new. And I think this applies to how we're integrating it. So if you want to do something new, and I mean, this is, you know, as a startup on the VC train. our daily bread and butter is to be the company that disrupts the incumbents and that innovates. So if we do something that an LLM can do, we can do, we'll do wrong things because they can just replicate it tomorrow. And it's a bit how we're trying to think of, you know, AI enhanced work that we're doing.

Michael Finocchiaro

Asparuh

Asparuh Stoyanov

Yeah, my view is similar. I don't think that ⁓ product development and this V-shaped standard form of product development will be fully autonomous by agents. First of all, because I wouldn't go on an airplane that was fully autonomously designed by agents, at least not until now. And I don't see it coming in the next couple of years. ⁓ I think that agents will. enhance and let engineers focus on this really engineering brain work that's interesting to every engineer who's in the field and automate everything else like extracting information out of documents, putting it in a presentation, ⁓ all these things. ⁓ I believe this is where the agents ⁓ and AI can shine to optimize the workflow. yeah, hopefully in the future, maybe we see major breakthroughs. ⁓ and things become even more autonomous, I still think engineers will be essential, humans will always be there to supervise.

Michael Finocchiaro

And ⁓ when a customer is using ⁓ EME or keyword, ⁓ where is the AI actually being used? Is it something part of the user interface? Are they interacting directly with an agent? Or is it part of the plumbing in the application? Where is the AI embedded in your work?

Dennis Just

So our products that we sell is the AI. So I mean, I'm not sure, Michael, if you tried the molding demo or injection molding that we released yesterday, right? So ultimately what we do there is we create an AI model that is capable to give instant physics feedback on any type of geometry process material parameters in the whole plastics injection molding value chain. So envision the AI taking the inputs that Naturally, a numerical solver would take and giving the output that a numerical solver would put out ultimately, you know, at the breadth and the depth of a whole engineering domain and in seconds. it's, it's in engineering, what we call engineer, but not in a sense of we want to replace the physical engineer. It's more like the little helper that does these things faster that for now the math, you know, took years to process. ⁓ doing it in seconds so that your iteration loops run way faster.

Asparuh Stoyanov

Yeah. And for us, the chatbots are not only the chatbots, the AIs are integrated into the platform in different ⁓ modules, different parts. We have bots that help our ⁓ engineers with creating reports. We have bots that help our users with ⁓ generating Python code if they want to quickly evaluate something that we don't provide as a no-call functionality. ⁓ Yeah, we generally don't allow the bots to do the engineers work in a sense that ⁓ we don't let the bots interact with the user's data, but they give the code to manipulate the user's data so that users can understand what's actually happening behind the scenes.

Michael Finocchiaro

So there's ⁓ some kind of an audit trail in terms of what actually happens.

Asparuh Stoyanov

Yeah, absolutely. You can trace everything. I think that's one very important thing with AI for engineering that you need to be able to understand what's happening and why is this decision taken. ⁓

Michael Finocchiaro

I think another aspect that might be ⁓ difficult to work people's head around is ⁓ the fact that AI is very probabilistic and engineering is extremely deterministic. So how do you guys have you guys put guardrails around it to make it more deterministic? Just a ton of MD files or maybe some other logic or maybe a committee of agents before a decision is taken. I'm wondering. how you guys get from a probabilistic model to a deterministic result. Because obviously simulation is deterministic,

Dennis Just

Yeah, yeah. And I think this is one of the biggest challenges, I think, from an architectural perspective that we encounter. the way how to solve it, or how we're solving is to create ways how to uncertainty or certainty quantify the results that the model ultimately gives the engineer ⁓ to then have a judgment call of does it hallucinate or not. And what you can do to get there is a kind of diffusion like approach. So, I mean, just to picture this a little bit. having an I model. So give the same input to the I model, have five outputs generated, compare the outputs to each other. If they are fairly different, and I'm oversimplifying, right? Then the certainty is low, which means that you have high uncertainty. If they are fairly similar to each other, then the certainty is high, and it's a result that you technically would want to recommend an engineer to use.

Michael Finocchiaro

Okay, interesting. How about you?

Asparuh Stoyanov

Similar to when you go to the doctor's office and you don't want to go to one doctor, you would like to go to a couple of doctors, get different opinions and see if they're like, yeah. For us, I mean, it's the same thing. If we talk about the machine learning models that users can train on their engineering data to do their predictions, as Dennis said, you can provide uncertainty quantification and measure this. And as an engineer, as a user,

Michael Finocchiaro

Save opinions.

Asparuh Stoyanov

If you see high uncertainty, means not really, I can't really trust this. ⁓ Yeah. And for the other bots that we have, the other agents that we have inside of the platform that help with actual data manipulation and ⁓ the workflows automation for them, we provide a large amount of contexts, including the knowledge base based on the customer's data. ⁓ So customers can. feed their engineering domain knowledge to the agents and the agents can use this information as a context to what they're doing. And of course you always need to provide where this information comes from, what was the source. Yeah, basically try to make it as traceable as possible.

Dennis Just

Exactly. Maybe adding to what Aspari said, I think one of the things that we're trying to work on a lot, and I think this accounts to any starter in the field, is to build trust on top of AI or applied AI in engineering. ⁓ And as we haven't had this GPT moment, there is still specific in large scale organizations and large industrial companies, there is reasonably so, I think a high amount of distance. towards trust towards these models. ⁓ But as the models get bigger, as they get more feature-rich in terms of being capable to have this uncertainty quantification, ⁓ I think if you ask the same question two years, I'm pretty sure that 50, 60 % of the people are actually using these models. It's just really a matter of time to get there.

Michael Finocchiaro

And in terms of the models, are you talking exclusively about the off-the-shelf commercial ones or the ⁓ open source ones? ⁓ If a customer is using Key Ward or EMI, are they able to bring their own because they already have Anthropic or Gemini or whatever? Or ⁓ are you also thinking of using some of the open source models with Elam Studio, Orlama or Microsoft Foundry in order to have a very trained local model that's not polluted with the with stuff that might be in Reddit or stuff like that.

Asparuh Stoyanov

Mm. Well, keyword is mostly focused on the ⁓ data part. we do provide the platform. In the platform, we provide the ability for users to choose which framework they want to use. And they can bring their own models. They can use PyTorch. They can use TensorFlow if they want to. They can use open source models ⁓ from hugging face. ⁓ Or third party providers could be even ⁓ MEI. ⁓ So yeah, we, Key Ward is more focused on providing the tools for mechanical engineers in an easy to use manner to use ⁓ models with their data.

Dennis Just

Yeah. And I think for us it's, you know, in the engineering domain, if you look at the process of engineering, there's no open model that helps you to do this in terms of building a product and getting an understanding of the physics. So you cannot use anything that is out there because it's very much built for context, text, audio and video understanding. So we build these models ourselves from start to the finish. that then you can use to put the geometry in, get your external aerodynamics out, or put your material in and get your filling pattern out for the injection molding, or put your, I don't know, car in and have it smash virtually against the wall. So they can get an understanding of how it crashes ultimately. So we are very much focused on building these own models that are capable to replicate true physical behaviors.

Michael Finocchiaro

So it's like ⁓ you guys are building your own, it's not really a large language model, it'd be more like a large simulation model or a large Emmy model. How does that work exactly?

Dennis Just

We call them large engineering model, but exactly. this is what we're doing. So these models are meant to be kind of your digital helper, your digital engineer for the engineering domain that they're built.

Michael Finocchiaro

How do you position yourself against ⁓ an engineering co-pilot like Leo AI, where Mo Aor, he basically took all these engineering textbooks and trained his thing on them and then put a sidecar on Onshape and Creo, and now the engineer has his assistant next to him saying, don't design it like that. Is it a similar idea, but for simulation or completely different?

Dennis Just

It's, think we're very orthogonal to that. So if you look at the way how you create a car, I mean, I would to, to, to generalize, I would more apply to the CD domain. Right. So this is where studying the textbooks and having an understanding of how you, you, create it ultimately, ⁓ from a design perspective is very helpful. ⁓ but then how do you evaluate what you've next to, you know, putting it into kind of a CAE tool that helps you to understand physics. And this is where we come in. I think naturally, and maybe I'm oversimplifying a little bit, you would create a car with what you described in terms of the shapes and the CAD and the materials. And then you put this into an ME model that gives you instant feedback of how it behaves, both from lift and drag coefficient, from crash testing perspective, from a thermal perspective within the engine. So this is what we're doing.

Michael Finocchiaro

So like ⁓ in the case of Key Ward, you guys are competing more against like orchestrators like Sinera or somebody like that or Simcenter ⁓ or maybe the Ansys Discovery platform, is sort of, no, not at all.

Asparuh Stoyanov

Mm. To be honest, not really. ⁓ We did used to develop our own models. We still do occasionally for some cases, but we are more of a platform. If you're familiar with ⁓ Databricks, ⁓ kind of pipeline data processing ⁓ orchestration, but very specifically ⁓ designed for engineering and manufacturing our workflow. So ⁓ To be honest, I'm not familiar if there is a similar tool out there on the market that does exactly what we do. ⁓ We kind of bind the data engineering, the data processing and the surrogate modeling into one platform. And yeah, that's what we do. ⁓ Scenera is more of a, as far as I'm aware, more of a orchestration tool for ⁓ multiple CAE. ⁓ that you can build a workflow and call these and execute these workflows, where Key Ward is more designed for mechanical engineers who want to do machine learning, want to do more data-driven analysis in a more automated way. And the platform can process any CAs over out there that produces data out there.

Michael Finocchiaro

Awesome. And I guess for ⁓ Emmy, the competition sits where exactly then? Because you guys are building our model. So it'd be closer to a neural concept or something like that, ⁓ or like Ansys Discovery, just a whole platform.

Dennis Just

And I think the newer concepts. No, so I think the new concepts and the physics acts of the world, they are more on the platform game of trying to replace answers for the engineer. Basically a two layers below that. So on a framework level, we are competitors to an Nvidia physics Nemo and Outer Desk NavPak. So I think this is very much the framework of the technology that helps you to create the AI. And on a level above on the model layer,

Michael Finocchiaro

Mm-hmm. Okay.

Dennis Just

We don't really have seen anyone yet. I think it would be more maybe the project Prometheus guys that are the closest to what we do there. ⁓ Exactly, but these are the ones. Exactly. So this is more the direction that we have.

Michael Finocchiaro

basis project. ⁓ So you think that OpenUSD will emerge as a standard that'll sit on top of that? Because I think that that's what it's going to be built on ultimately, ⁓ My understanding of Prometheus is sort of the multiverse, right? It's a physics-informed multiverse. And I think the standard for that is OpenUSD. That's the standard that ⁓ NVIDIA and PDC and Siemens and ⁓ DS have all signed on to as how do you...

Dennis Just

Say it again, Michael. you

Michael Finocchiaro

do physics, how do you connect the physics, the IOT stuff to create digital twins, right? And then have a closed loop to get the feedback.

Dennis Just

I don't, so I mean, we know quite some peeps within there and I think what they're trying to crack the foundational intelligence layer for engineering. So I think they're even one layer below of what you were describing. And from what we can judge from the outset, I think they very much as well go the route of data driven models. They're not at all, at least from what we know on the physics informed, a side of what they want to create. And maybe this is tied at the beginning more towards the Bezos type companies like Blue Origin, et cetera. They really have to go into external error or mechanical stresses specifically for space. I think that's the reason why when we develop what we develop and then challenge it to the market, they are usually the ones that we see the most.

Michael Finocchiaro

Interesting. Sorry, those are sort of ⁓ unscripted ones, but I just find it so interesting that we're really hitting a lot of the edges where nobody's really gone before. We're sort of the Star Trek already, right? ⁓ In terms of like the future, so we've been four years into the AI revolution now, right? We've seen ⁓ Chappie T3 and 4 and 5, and now we've seen Claude Code and Claude Botts and that's Molt Book, whatever that is. ⁓

Dennis Just

you Thank you. ⁓

Michael Finocchiaro

Are you guys as skeptical or bullish as you were three, four years ago? And where do you see it going? I guess, Dennis, you already said you're waiting for engineering's GPT moment. Does that happen in 2026? Does it happen in 2027?

Dennis Just

I mean, so I think that that happens in 2026, but it happens in domains. So what is different from language to engineering is that it doesn't make sense to build this one model who does it all, right? Because

Michael Finocchiaro

Dennis?

Dennis Just

the physics, no matter where you start being at the generative design part or you're more on the CAE part of solving structural mechanics, whatever you solve for electromagnetics, is very different from domain to domain. And this is why I think we have very much an approach of focusing on these specific use cases, but then go as generalized as possible. So we will see the first foundation models in engineering this year.

Michael Finocchiaro

Wow, that's pretty bullish. ⁓ Asper, do you agree?

Asparuh Stoyanov

To be honest, I'm maybe a little bit more skeptical about how AI develops. My opinion hasn't changed since GPT is out. I just think that there is a lot of hype, definitely from what I see from being in the industry and talking also with clients in the industry. I agree with Dennis that There will not, engineering is fundamentally different in the sense that there will not be one universal model that can do everything. I think that partially the reason for this is that data for engineering is not as abundant as text online and images online. And it's just difficult to train such large model. So I do agree. ⁓ Future is probably more focused foundational models for specific ⁓ problems. we will get to this, to a model that really can generalize well for a case. It be great for this year, let's see. But I just think it will take a couple more years.

Michael Finocchiaro

But do you guys also think that we're going to see the design and simulation and design and manufacturing that time compressed? Because it seems to me like you've got all this ⁓ work on simulation, but we've also made a lot of advances in CNC programming and tooling and FIS access machines. And those things, if we're going to accelerate time to market, those things have to happen almost simultaneously. They shouldn't be. You know, I have an engineering thing. give it to the simulation guy, wait a month. He comes back with something and then I pass it over the wall to manufacturing and they mess with it and they come back to me six months later saying, actually, that doesn't work. go, this is not efficient, right? And, and I keep talking to these companies and there's this huge wall, right? Between manufacturing and engineering. Shouldn't AI be destroying that wall and just blinging us all together? I mean, I know you guys are more on the simulation side, but you've mentioned manufacturing a couple of times. There must be a lot. of fundamental physics, there is a fundamental physics in terms of machining and milling and operations like that. So what do you guys think? Do you think also we'll see a much more compressed timeframe where the engineer has to be multidiscipline. He's got to be able to do, or she has to be able to do both design, but also think about the physics and the manufacturing, the simulation. Sorry, I'm going on a long time to the question, but I think you see where I'm going. You want to try that one, Asprey, and then Dennis can jump in.

Asparuh Stoyanov

⁓ Yeah, so ⁓ I do think that AI will have impact in terms of increasing the speed at which products are developed and reduce the ⁓ time it takes to ⁓ develop design and also to manufacture a product. ⁓ think partially this bottleneck comes also from the way companies work. ⁓ I find the engineering field not as data-driven as financial. ⁓ in the financial sector, like machine learning AI, data driven tools have been used for a long time there. And in engineering, I people are just starting to change ⁓ their approach to how they analyze data. And I think AI can support them in this transition there. ⁓ I read, like some time ago, I read an interesting report where, ⁓ yeah, basically with LLMs, with charge GPT, you are able to do ⁓ 10 times more. ⁓ cold, for example, spending the time to actually fix the cold after when it fails in production can be also counterproductive. So it's important how AI is also being applied by the companies to really boost their efficiency.

Michael Finocchiaro

Dennis, you got some comments?

Asparuh Stoyanov

Yeah.

Dennis Just

I think so. Our view is, I think ultimately you would need to, and I think it's very parallel to what you've seen in LLM. think you first want to reach a level of intelligence, kind of the foundational layer, being it in generative design, like being it in physics, physics simulation, being it in PLM. And what will happen afterwards is that, and I think this is very much what is in the likes of trying to do and I think we'll be successful doing so is you try to then create this in an agentic way so that these kind of elements can talk to each other and you have references, right? And then kind of the whole automation and being capable to run a workflow from A to Z basically starts. But the stage that we're in at the moment is that we're trying to break through this foundational layer of intelligence of one domain or of each domain to then be capable to replicate this across the whole.

Michael Finocchiaro

Interesting. so related to that, the younger, look in the demographics of people watching this podcast, I do have a lot of entry level people. And so there's a lot of anxiety on how AI is going to take their jobs. What do you think they should be focusing on, this younger generation of engineers that we, and why should they work for startups instead of for the big three? That's another one I keep forgetting to ask. Cause I mean, obviously you want to get the best brains. working for your startups and that they don't go to Nvidia and Daso and Google and so maybe I turned it that away like once they learn it and why would they go work for you guys and take that risk rather than working for a safe boring place like Facebook, you

Dennis Just

I think you're underestimating how good startups are paying, to be honest.

Asparuh Stoyanov

Who wants to start? Shall I start? Okay. Well.

Michael Finocchiaro

Good for it. Sound problem.

Dennis Just

Go ahead as per-

Asparuh Stoyanov

⁓ Yeah, I mean, you say it, Michał, boring job at Facebook or Meta. ⁓ I usually say if you want to learn politics, go to big companies. If you want to learn ⁓ the domain and the skill set, go to a startup and you just be encountering so many problems every day. You just learn to think fast and how to solve every problem that comes your way. I personally have always preferred this way of ⁓ working. And I think for young engineers, young people coming out of university with AI around, that's the way that you are going to get ⁓ valuable experience that will make you irreplaceable by AI. I do believe you need to adapt and use AI tools. ⁓ I do, I find that young people coming out of universities really do struggle right now because ⁓ I personally don't need five juniors right now to assist me with my work because I'm using AI. ⁓ But still, ⁓ there is a need for juniors and they're going to learn more when they ⁓ work in a startup. It's just fast-paced environment.

Michael Finocchiaro

Dennis.

Dennis Just

Yeah, I mean, ultimately, I think, you know, I don't think that the AI and we see this in any other domain will replace people fully, right? I think the way we work will change and the velocity of what we input and output in our work will increase with the help of AI. But as long as you stay curious, I wouldn't worry, to be honest. If you get lazy, it might be different. But I think that really depends on your personality. ⁓ that you have. think generally speaking, I mean, we are super lucky of being in a phase where, you know, if you have great talent and great talent kind of recognizes the speed of what is happening at the moment, ⁓ they want to work in startups to learn and to learn fast, to have high velocity. ⁓ So we're extremely happy to have the people that we have in the team.

Michael Finocchiaro

We have one question coming in in the comments from Victor Rosa. says, thanks to both of you for your insights. What is your assessment of the current landscape of benchmarks around AI physics models? What are the challenges towards standardization of the models coming from the open source versus commercial communities?

Dennis Just

Yeah, happy to take this one. Yeah, as per go. I mean, it's, I think for us, honestly, I think it's great to have these benchmarks to have a way how to measure accuracy and model capabilities. But what we do see with, you know, being a driver, being a driver net, being at the new kind of error benchmark from Nvidia, that are all great.

Asparuh Stoyanov

Let me see, you want to go first?

Michael Finocchiaro

because it's your question, think Dennis.

Asparuh Stoyanov

No, no, I think it's your motto on your slide.

Michael Finocchiaro

It's.

Dennis Just

benchmarks in data sets is that they're very narrow in terms of applicability. So they are great to, ⁓ I'd say, benchmark surrogating capabilities, but they're not good to benchmark foundation model capabilities. So I would love to see broader, bigger data sets that actually cover full domains rather than very narrow ⁓ data sets that are more kind of toy problems, if you want to say it this way, that really help you to benchmark. you know, true capabilities rather than little features ⁓ of certain models, AI architectures, et cetera.

Michael Finocchiaro

Thanks. Anything to add, Asprey, or we move on?

Asparuh Stoyanov

No, I fully agree with Dennis. The benchmarks are good as orientation, but they're not really objective always. ⁓ A larger database, a larger amount of engineering data being open source can really ⁓ help creating better models. ⁓ Architecture plays a bigger role, but I data plays a larger role for the performance of the models.

Michael Finocchiaro

My data is always the central linchpin to all of this stuff, I think. So let's switch gears to digital maturity and enterprises. So when I created the podcast, it's because a friend of mine asked me, said I should do it on AI, but I thought, well, nobody wants to hear me just say I'm at AI and no customer's going to come on and tell us what they're doing with AI. But talking to you guys who are customers, and probably you can give me a better feel for what's actually out there.

Dennis Just

Maybe.

Michael Finocchiaro

When I look at AI, think of a spectrum from one to five. think of like one is like, I'm still using email for most of my collaboration. I'm still using Excel for 99 % of my bombs and just about everything is Excel and email. And five would be fully autonomous, agentic, ⁓ adaptive digital twins and nobody's at five, right? I mean, almost nobody. ⁓ So in your experience and the companies that are buying Emmy and ⁓ Key Ward, The companies are working with closer to closer to one or two. Are they anybody at three? I how do you assess the without naming names? But in general, where do you see your customers in terms of their maturity?

Dennis Just

Go ahead, Asparu.

Michael Finocchiaro

Dennis? His Vizio might be frozen. Asprey, you want to take that?

Asparuh Stoyanov

Yeah, I'll take it. Yeah, so maybe until two years ago, would say the majority of the companies we were talking to were at ⁓ one or two. very in the beginning, some of them were just starting to get on the AI wagon, ⁓ mostly because of hype, based on LinkedIn, ⁓ social media. And... What made me an impression is that from last year, from the conferences that we were attending, ⁓ I see that more and more companies have already done some initial projects. They're a little bit more ⁓ aware of the challenges with AI models, such as everything surrounding getting your data, preparing your data. And I still encounter companies that still haven't started, but let's... Let's say that those are right now maybe 10, 20 % of the companies we interact with and 80 % have already started with some POCs or they're at stage two, level two or level three.

Michael Finocchiaro

So they're passing two and heading to three basically.

Asparuh Stoyanov

Yes, yeah, very few are at stable level 3 going to level 4.

Michael Finocchiaro

Okay, is that your same experience Dennis?

Dennis Just

Yeah, sorry for the internet. have no idea who is, is, ⁓ I don't know, cutting the cables here. ⁓ so I think as, as we are more on the fundamental level, I think I would rate our customer kind of projects more on three, four, and maybe scratching the five at the moment. So think the most advanced AI models we're building are really models for controlling large scale engineering processes and systems.

Michael Finocchiaro

It's okay. Hahaha.

Dennis Just

to optimize them, so ultimately to use kind of this real-time inference to inform, you know, I think the network models to steer large scale products, being it, I don't know, from chemical systems, kind of the rotary kilns and the processes that you have there, towards kind of grid capabilities that you have. So these models exist, but they are very rare and they're not common, I would say at the moment. So we're definitely moving towards the four five, but in very specific cases where you have enough data and where you have enough value for the money that you need to train these large scale models.

Michael Finocchiaro

I suppose that means that there's just part of the company that's that advanced because I don't know if there's any company that the entire company is that far along, right?

Asparuh Stoyanov

I believe that Dennis was referring to the maturity of the technology. Yeah.

Michael Finocchiaro

Yeah, and I'm talking about the companies, your customers' maturity. Thanks, Aspara.

Dennis Just

so sorry, then maybe missing this one. So I mean, on the customer side, I'm with you. So I think this definitely ⁓ is not far down the road. I think we have maybe to make a bit of an example. think there are very few companies who actually approach us with a problem that they want to solve with AI. They mostly approach us with ⁓ the intent to include AI in their processes, which can be very painful.

Michael Finocchiaro

Hmm.

Dennis Just

because you become the McKinsey for them. ⁓ And they did rather pay the consultants, but yeah, no, fully, fully agree.

Michael Finocchiaro

Yeah. So my thesis is that putting in AI powered awesome software like Key Ward and EMI is a catalyst. It's a way of moving the needle to the right. ⁓ And that the big three are all kind of floundering and they're absolutely not delivering state of the art bleeding edge AI functionality today. ⁓ So how, in your experience, have you seen that, ⁓ have you seen a customer have an aha moment or an epiphany when they were using Key Ward or EMI and said, Holy cow, if I break my data silos and I have data governance, I could actually get so much more and actually really innovate and digitally transform my enterprise into something that's far more powerful. So have you actually seen that in action in your experience with customers?

Asparuh Stoyanov

⁓ Yes, that's actually one of the most satisfying parts. I'm personally constantly in touch with the customers. I'm usually onboarding them on the platform and teaching how to use it. And for many of the mechanical engineers, it's a new way of looking at data. Usually in the typical engineering workflows, engineers would do one simulation or one test. They'll look at the measurements. Maybe they'll put it in a PowerPoint report, but they'll forget it. And what we... ⁓ Provide is a way to collect all this data, put it in one place, give a full overview, analyze it more kind of like a big data, ⁓ look more at trends, spot patterns, identify ⁓ hidden knowledge gaps that you haven't talked before. And usually when we have a trial with a company starting from one user, ⁓ very soon after the people start interacting with the platform. So then there are 10 other use cases from other teams that can be ⁓ explored. And yeah, that's it's satisfying moment.

Michael Finocchiaro

So my thesis is validated. How about on your side,

Dennis Just

Same thing. we actually, think most of the customers we have had failed projects with others before on the problems they're trying to solve. So it's kind of the classic thing of ⁓ over promising under delivering. ⁓ So this is usually from a technology perspective, how we then come in, you know, try the challenges, break through the challenges and then kind of have happy customers afterwards.

Michael Finocchiaro

Okay. So really, ⁓ do you feel that just the fact that you guys are so AI native, that's what's giving you such a wedge against the Daso, against Simulia and Ansys, ⁓ Simcenter, all that stuff. ⁓ It's just your agility and, how do you compete against the big guys? it basically on that agility and that being close to the customer as well, understanding really their pains?

Dennis Just

I think for us it's more on a capability side of things, right? Because I think it's just very different what the big docs that have reasonability to sell their seats kind of put to the market. The way how AI will impact this whole thing is so different towards what they can offer and what they're selling. But because how do you want to sell an AI model or AI capabilities on a seat-based license? Like this doesn't make any sense. So you have to fundamentally rethink the way how you price, how you productize. And given the margins, the reach in the market web that all of they have, I think, are not incentivized to change towards this kind of structure, which would kind of naturally or inherently is an issue and allows us being deep learning native or AI native, however you want to describe it. with the capabilities and with the speed that we have to just hardly compete.

Michael Finocchiaro

Interesting. Same experience on your side.

Asparuh Stoyanov

Yeah, for us, we definitely have the speed advantage. We work and make decisions very fast. We have iterated for the last four years to find a product that really meets users' needs. ⁓ And we are just capable of If I have to give concrete examples, I was at a conference last year and one of the company's major software vendors stated their roadmap to develop a specific agent in their platform that we actually developed two weeks earlier. And it took us approximately two weeks from prototype to being in production. And in their roadmap, this was announced for something that will come next year. So the speed is much higher.

Michael Finocchiaro

Two weeks. Wow.

Asparuh Stoyanov

⁓ Yeah, I think that's the main advantage towards the big guys.

Michael Finocchiaro

That's really fast. Amy, how about you guys over at dinner?

Dennis Just

So I mean, I think that the good thing is tying towards in terms of the technology that we develop, are very orthogonal to the big dogs. ⁓ So we see the same, think, in terms we have a 30 people organization that, you know, 25 people are developers and research engineers. And we're trying to interact super fast. And I think that's the reason why we're capable to shoot out all these kind of foundational things with a very slim org being very efficient and effective with what we do. compared to multiple layers of management, old structures, ⁓ products that you have to maintain. We basically have a head start by building from scratch.

Michael Finocchiaro

So like for, for, for any, like from a meeting the customer to deployment, is it like what Aspro is saying is something that a matter of weeks rather than a matter of months or years?

Asparuh Stoyanov

So.

Dennis Just

I think it's more a of month for us at least because the, let's say the time to generate the data, train these models, implement these models obviously take time. But if an incumbent would do that, I think it would take them like years to get there.

Michael Finocchiaro

Sorry, Asprey, what were you gonna say?

Asparuh Stoyanov

In our case, ⁓ it's instant because we provide a platform that if we give you credentials now, you can instantly log in. And depending on your use case, if we are talking about a 3D deep learning use case that they want to train, it can take a couple of weeks until they have ⁓ a trained initial model. ⁓ If we are talking about time series models or tablet data models, ⁓ we have delivered value within a day, within two days even. companies that they already have a model that gives them good results.

Michael Finocchiaro

It's really incredible because the big three, you could be waiting a considerably longer amount of time for less of a result, right?

Asparuh Stoyanov

yeah, absolutely.

Dennis Just

and have two zeros at the price tag.

Michael Finocchiaro

Yeah. Well, that's fantastic. That ⁓ was really awesome. And it was interesting to see, Dennis, you're working more on the fundamental layer. And Asprey, you're working on making sure, keeping the data clean between the process. Really complimentary solutions, actually. ⁓ Did you guys have a good time, too? ⁓

Asparuh Stoyanov

Exactly. Yeah.

Dennis Just

Yeah, thanks for taking the time

Asparuh Stoyanov

Absolutely.

Dennis Just

and grilling us.

Michael Finocchiaro

Where can we see you guys in the next couple months? Are you guys coming to Threaded and Warwick, for example, the best startup conference in the tech world? ⁓

Asparuh Stoyanov

I heard it's Florida, early April.

Michael Finocchiaro

Yeah, there's Miami too, because I did Miami and Threated War. But seriously, are you guys going to be at CDFAM? Are going to get people to come and meet you at any conferences in the next couple of months? Other than my conferences, of course.

Dennis Just

I we'll obviously be there. No, no. I mean, I think for us, we are more on the deep learning kind of conferences. So the New Europe's ICML type of conferences at the moment. I mean, feel free to come by and otherwise always in between Berlin, ⁓ Paris and Vienna and Linz. So happy to meet for coffee.

Michael Finocchiaro

Okay. Thank you. Yeah. or beer like we did with keyword. How about you, Asprey?

Dennis Just

or beer or schnitzel.

Asparuh Stoyanov

For us, are mostly actually focused on the engineering conferences. We'll be participating in NAAFAMS, ⁓ believe ASSESS, then a couple of other automotive and ⁓ manufacturing conferences we're going to this year. But to be honest, I don't know all the names. I will be at NAAFAMS and it's probably in Florida with Umica. And otherwise, we are also based in Berlin. I'm currently in Berlin, but we do have also

Michael Finocchiaro

Nice. great set.

Asparuh Stoyanov

offices in Amsterdam and Paris and in the US and Colorado. So yeah, also open for a beer or a drink, tennis if you're in Berlin. I would love to have a coffee with you and talk how we can collaborate.

Michael Finocchiaro

I

Dennis Just

will do.

Michael Finocchiaro

didn't know you were there. Otherwise I would have come when I came for Radiohead back in December, which is a fantastic show at the Uber arena. ⁓ Well, so to the audience, thank you for joining. We had about 15, 20 people. ⁓ You guys feel free to reach out to Asperon and Dennis if you have about EMMI AI or keyword.

Asparuh Stoyanov

⁓ Yeah.

Michael Finocchiaro

And we'll see you on the next podcast. Tomorrow I have a Manukai and productive machine. So we'll be switching back to the manufacturing side after having a little tour of ⁓ the simulation side. So thank you very much. And we'll talk to you on the next AI Across the Product Lifecycle podcast. Bye bye.

Dennis Just

Thank you.

Michael Finocchiaro

Okay.