Closing the Loop: AI in Manufacturing Workflows — with Lambda Function and up2parts

Episode Summary

The episode "Closing the Loop: AI in Manufacturing Workflows — with Lambda Function and up2parts" delves into the integration of artificial intelligence (AI) within manufacturing workflows, featuring Tanmay Aggarwal from Lambda Function and Marco Bauer from up2parts. Both companies leverage AI to enhance precision manufacturing processes. Lambda Function focuses on transforming machine-generated data into actionable insights using AI and machine learning techniques, aiming to revolutionize how manufacturers process and utilize data from CNC machines. On the other hand, up2parts specializes in automating quoting and machining part proposals for CAM projects, integrating AI to streamline these tasks and improve efficiency.

Key technical and strategic insights discussed include the importance of setting realistic expectations regarding AI's capabilities and limitations, as well as the necessity of redefining one’s role within manufacturing workflows. Both guests emphasize that while AI can automate routine tasks, its true value lies in enabling engineers to tackle complex problems by leveraging AI tools to augment their work rather than replace it entirely. They advise professionals to focus on identifying meaningful challenges and using AI as a tool to enhance productivity and innovation.

For PLM and engineering professionals, the episode underscores the potential of AI to transform manufacturing workflows by automating mundane tasks and empowering engineers to concentrate on more strategic problem-solving. The key takeaway is to embrace AI as an ally in the pursuit of efficiency and innovation while maintaining a critical perspective on its applications and limitations.

---

Full Transcript

Michael Finocchiaro

Okay, I think we're live. This is ⁓ Michael Finnecaro on the AI Across the Product Lifecycle podcast. And I'm joined today by Tanmay Agarwal. I hope I said that right, Tanmay, of Lambda Function. And by Marco Bauer of Up to Bars. I'm very happy to see you guys. How are you guys doing today? Good? Great. So... ⁓

Tanmay Aggarwal

Want to live?

Marco

Very good, thank you.

Michael Finocchiaro

We wanted to first I wanted you guys being to introduce yourselves and what your companies do before we jump into the more AI ⁓ questions. So ⁓ do you want to lead us off Tanmay on what you do and you your parkour and everything?

Tanmay Aggarwal

Yeah, absolutely. Thanks, Michael, for having me and Marco. Great to see you again. So my name is Tanmay Agarwal. I'm the founder and CEO at Lambda Function. We're based out of San Francisco. We started in 2020 ⁓ with the idea of looking at how AI ML might impact the world of precision manufacturing. ⁓ Fast forward five years later, I think all those teams have converged really well and we're going to see kind of a lot of change in the market, of course. So excited to be here and excited to chat more about what we're seeing here. Marco.

Marco

Thank you, my name is Marco Bauer, I'm the founder and CEO of OptoParts. We are located in Germany, in Bavaria. ⁓ Our company started also 2020, but the history is going back to 2010, when we started a manufacturing company. So ⁓ that was the groundwork of OptoParts. Then we started to develop a software, first for our own demands to automate the coding. because that was the biggest pain point in our manufacturing company to speed up this process. ⁓ now after five years later, ⁓ we are also ⁓ doing AI automation for ⁓ quoting machining parts and also do proposals for CAM projects and transfer this information to existing CAM systems.

Michael Finocchiaro

really cool. ⁓ So I guess both of you have been pretty bullish on AI from the beginning. So Marco, I mean, probably you were doing even AI before the JinAI LLM thing, but what was your feeling when all this stuff came out, when they opened AI kind of opened the floodgates, if you will?

Marco

⁓ I think in the beginning I was on the one hand quite skeptical, but on the other hand I also see the chances for AI. ⁓ I think the most important thing where I was concerned about is ⁓ about the right expectations on AI and about the right usage of AI. So I think it's a technology where you can, ⁓ when you have ⁓ not the wrong expectations, when you have wrong expectations or when you use it in different way you can do good things and also bad things with AI. So ⁓ I was in the beginning more skeptical I think.

Michael Finocchiaro

How about you, Timmy?

Tanmay Aggarwal

Yeah, it's interesting. I'm reflecting on like back in 2020 when like I remember playing around with GANs in 2020. And at that time GANs were being used for, ⁓ you know, art work, like being able to recreate, let's say, a historical image of an art using a GAN model. It was quite impressive at that time, right? And then when LLMs came around in 2022, 23 time period, ⁓ I think it's actually pretty amazing ⁓ what these models have been able to do. ⁓ with traditional kind of natural language processing, which is essentially what these are doing. So when they started, think like my biggest skepticism was around what could these models do for non natural language use cases. And I think what we do at Lambda Function, right, a lot of what we do is taking the data out of CNC machines and trying to create intelligence out of machine generated data, essentially. ⁓ And so that question really has, and it continues to persist in my mind, which is at what point in time will these models get to the point where they can start actually making sense out of machine generated data? So that's, think, where kind of where my question has always been around. But I think it's incredible to just see what these models can already do. And I think that just goes to show the power that these may have.

Michael Finocchiaro

Absolutely. And we'll get into that a little more later in terms of that. ⁓ But before that, since you guys have both ⁓ created software platforms, there's a lot ⁓ of hype around vibe coding. But in fact, I think the fundamental job of every developer in the planet has been fundamentally changed since the advent of Cursor and all these other tools. ⁓ How on a daily basis are your developers, Tanmay, using a... these tools to enhance their experience as a developer? And obviously, how is that impacting a Lambda function as an app in terms of the development process?

Tanmay Aggarwal

Yeah, I think it's interesting, very timely, because just last week we had an internal all hands talking about similar topics. And I think the answer to that, Michael, I would say is like there's one more of a citizen developer perspective on adopting AI, and then there's a more of an organizational view on adopting AI for improved software development. ⁓ I think over the last few years, we've definitely seen individuals within our team who've experimented with. ⁓ tools to accelerate their workflows. And that's been great, I think. But now I think the biggest transformation that's happening within Lambda Function is we are taking it as an organizational priority, which is, you know, how do we transform our entire SDLC cycle from the traditional way of doing software development to now a more AI augmented way of doing software development? And I think that that is a huge transformation. ⁓ I think the traditional approach of ⁓ the classic kind of scrum models and themes is changing very dramatically. ⁓ I'm hearing more and more of this concept of team of one. That's a concept that I'm myself introducing within the organization, which is that each developer should think of what would they do if they were a team of one? How could they be a full stack?

Michael Finocchiaro

Agile.

Tanmay Aggarwal

⁓ end to end, you know, with leveraging kind of AI co-pilots to be the pending, let's say, skill sets that maybe they were missing as an individual. And I think that's least the direction we are going. And I think that's, ⁓ you know, for us, it's an existential question. I think as an AI company, we need to adopt some of this ⁓ and explore that and then define what that means for the future. If we ever think of competing in this world going forward.

Michael Finocchiaro

Yeah, I think ⁓ on my previous podcast, one of the ⁓ founders said he could tell who was using it who was not in terms of how productive they were. The ones that were ⁓ using AI were far more productive than the ones that were not. ⁓ We seem to have lost Marco. ⁓ I hope he'll be back. ⁓ Are you guys ⁓ using AI primarily when you're developing? you using like Cursor and ⁓ or are you using this on the side? mean, how, how are you using it?

Tanmay Aggarwal

I think honestly, we don't yet have a formal mapped out, here are the tools that we are using versus not using. think we are kind of right now been a much more of an expiratory approach where we want to let different team members explore different tools that are out there. But I kind of see it from the lens of. Where can we kind of see the highest ROI with the lowest amount of change management associated with it? But for example, like code reviews is something that I've been seeing a lot of AI being used for. And I think that's an interesting use case for it. think how can we further strengthen our QA testing processes and other interesting use cases. So I think I look at it from that lens of what's the bottleneck? Where can we see this as being an augmentation layer in the near term?

Michael Finocchiaro

Yeah, I was thinking too, like I see a lot, or I've heard a lot of people using it for product management, like creating PRDs and things like that makes it a lot easier. you're back Marco, thank goodness.

Tanmay Aggarwal

when Yeah, exactly. Yeah, Michael's back.

Marco

Yes, sorry. Power shortage in Germany.

Michael Finocchiaro

thought maybe it was another CloudFlare error like I had an hour ago. How are you? So we were, ⁓ Ted May was just talking about how ⁓ at Lambda function they were using it. Okay, so how about adapt to parts? hear about that.

Marco

Maybe. Yeah, I heard it. ⁓ We are using LLM tools in different phases by the development. are using GitHub Copilot for sure for the senior developers. That's important that we decided that only senior level developers are using LLM tools because we think it's important that the junior developers, ⁓ yeah, they don't...

Michael Finocchiaro

They need to have better chops. ⁓

Marco

Yes, and ⁓ only rely on LLM, ⁓ it doesn't improve the quality of the code. Then we're using ⁓ LLM systems for quality assurance in the end, ⁓ by checking the code. ⁓ And I think for up to parts, especially in the sales, it was very helpful to integrate LLM tools. ⁓ For example, in the past we did a lot of ⁓ ... screen costs and anything else by recording with native speakers ⁓ with high costs and today we use LLM tools to record the video in an LLM generated English and then we ⁓ transfer this screen costs in no time into Chinese, into Japanese ⁓ and the quality is quite impressive. And ⁓ now from the product perspective ⁓ we are working also to

Michael Finocchiaro

Yeah.

Marco

to use LLM models to communicate with the users, because we see that, especially when you have very complex technical issues, you build trust when you have a co-pilot interaction where a LLM model is communicating with the user ⁓ in normal words and not just have a list of features or a list of tools. ⁓ I think the user is feeling more trust into the system when their LLM is explaining. this is a feature, there is a pocket, ⁓ we decided to choose this flat end mill, because the tool lifetime is good, the speeds and feeds are good and so on. And for that LLM is ⁓ quite an impressive technology.

Michael Finocchiaro

⁓ You guys are both just leveraging it. ⁓ Well, let me just go into the second question, because that's where I was going with that. So when I'm using ⁓ Lambda function or up to parts, am I interacting directly with an AI or is there more AI underneath the covers? Where is it actually implemented in your system,

Marco

⁓ I think at the moment it's more ⁓ underneath. ⁓ So we have AI tools in the background, ⁓ like simple services, like a process predictor, where we decide if it's a lathe part or a milled part. ⁓ And ⁓ also in the back end, where we're doing data analysis, the feature recognition. But the UI at the moment is not really like a copilot ⁓ interaction. It's like a standard UI. and it's more underneath.

Tanmay Aggarwal

Yeah, I would say very similar, Michael. are, see, a couple of thoughts there. One, think from our experience today, think customers probably don't care much about the technology. They really just care about what can the technology do for me in terms of what business impact is going to drive. So I'm a pretty strong believer of kind of keeping the technology.

Michael Finocchiaro

for lambda function.

Tanmay Aggarwal

behind the covers. It really shouldn't matter to the user. ⁓ independent of that, think currently within Lambda as well, we see ⁓ all of the utilization of AI currently still being behind the scenes powering the various use cases that we're going after. I do think there's a very strong ⁓ argument to be made to have ⁓ a UI layer where the end user does interact with that as well. But I think We're just not there yet in terms of at least Lambda functions journey of building that into the product ⁓ priority roadmap.

Michael Finocchiaro

But are there things that you think that you would never use AI for that you'd like, no, that I'll never delegate, that's got to be human, or you think it's all a matter of time that the LLMs get mature enough to handle just about everything?

Tanmay Aggarwal

So ⁓ from my lens one, think ⁓ LLM is one tool in the toolkit. Even within Lambda, actually, lot of our AI application has been traditional ML ⁓ in various forms. Again, it's also because we started prior to the whole LLM wave kicking in. And today, we think of LLM as another tool which can help augment what we're doing in some ways, which is great. ⁓ And yes, do think that I think, ⁓ you know, I think eventually I'm starting to be a big believer in the whole agent-tech AI ⁓ conversation, Michael. So I do think that that framework fits very well with our original hypothesis behind unlocking autonomy in the manufacturing domain. And you think about kind of the five levels of autonomy and you think about, you know, currently the world being at kind of more of an AI assistant. But we are starting to see more and more customers wanting to get into level two, level three type of autonomy, of high risk autonomy. So you kind of need these agentic solutions that can allow for that degree of autonomy. So I do think that eventually maybe we'll get to a point where certain workflows can be fully autonomous. But I think it'll be very workflow specific.

Michael Finocchiaro

Yeah, that's a good, actually I was using, when I've been talking about it, I've been using Ayora Berry of PTC's thing of advise, assist, automate. It's a three level, it advises the copilot just telling you stuff, assist you do want, like here's a bunch of options and which one you want me to do and then it does it and then automate, you give it a workflow. So I can concur with that idea of having several stages. How about you Marco, is it the same kind of concept for up to parts?

Marco

⁓ Yeah, think so. ⁓ I think manufacturing is too complex to think to do everything with AI or with LLM models. It's also a part of the tool set. But yeah, I think in the end it's important to use different tools and to try to automate what's possible and what really brings a user benefit for saving time. but not always to try to automate everything, ⁓ to put a lot of efforts in developing some stuff ⁓ and then you save in the end maybe 2 % of your whole process time. ⁓ For this, manufacturing is definitely too complex that you can do everything in short time.

Michael Finocchiaro

⁓ That reminds me, like when I interviewed Karan Talati of ⁓ First Resonance, we were talking about ⁓ that in engineering in general, we like deterministic things. And the thing that bothers us about AI, particularly LLMs is they're probabilistic. And then Karan very brilliantly said, but, know, manufacturing is pretty probabilistic too. It's the head engineer might not show up. He might be sick. A guy could make an error. The guy could... cut his finger off in a mill. mean, anything can happen, right? ⁓ But besides that, mean, how are you guys putting in those guard rails ⁓ so that you don't get those hallucinations where you're asking about a miller and he gives you an answer for ⁓ a drill? mean, what kind of guard rules are you putting in place to avoid this kind of hallucination?

Tanmay Aggarwal

I can go first, Marco. ⁓ Yeah, I think it's a great question there, Michael. think from our lens, I think that's where the subject matter expertise comes in. And I think the domain... Yeah.

Michael Finocchiaro

the human supervision or like training the AI on the expertise.

Tanmay Aggarwal

Various aspects to it, honestly, Michael. I ⁓ think both the domain Marco and I, we play in is a very complex enough field where this is not traditional information you just find on the internet, right? And there's a fair bit of subject matter expertise that people have acquired over the decades of doing this trade. And at least at Lambda function, a lot of what we've had to do is kind of painstakingly taken the time to transform some of that ⁓ intellectual know-how, human know-how, human capital, into now, ⁓ into code that we can feed into an LLM or alternative ways of applying AI techniques to it. And that, for us, creates guardrails. Those guardrails are very, very important. I think in differentiating what a traditional vanilla LLM could do in our subject matter versus an LLM powered with Lambda functions proprietary own data sets could do. ⁓ Similarly, think introducing more physics-based ⁓ machine learning techniques can further guardrail some of the output that otherwise theoretically dumb AI models will not know. to do with. that's my two cents on it. Marco, can you do here?

Marco

⁓ Yeah, so in our solution the interaction with the user is quite deterministic. ⁓ we have ⁓ like a quality assurance level ⁓ between the AI and the user interface. Because when it comes to manufacturing I think it's very important that you don't have a probability. Because ⁓ our machine tools are quite expensive and ⁓ it's important not to have a not rely on probability. ⁓ So think in manufacturing it's important that 7 plus 5 is equal to 12 and not probably 12. ⁓ And so we are using the AI in the background more for ⁓ analyzing historic data of the customer and then we recognize patterns within the data of the customer. then the result out of this ⁓ analyzation is like rule sets, where we see there are different routes. in a manufacturing company and then often we use these routes to discuss these routes with the customers because sometimes you see in old data ⁓ some routes without any sense and then you discuss with the customer and then you realize ⁓ at that time we created this data this one milling machine was under maintenance and that was the reason why we took another route. ⁓ please delete this route because that's in our standard route. And then we give this quality controlled system or rule set into the application and that's what the user is working with. And so I think the user experience is quite deterministic ⁓ and I think that's very important to build up trust.

Michael Finocchiaro

So are there, so are there, so is the idea also to find like smaller workflows, like maybe there's things that are less performance critical that couldn't damage the machine and those things you can automate, like maybe some of the cleanup tasks or ⁓ administrative tasks and then some other tasks you always have to have the human in the loop, like you were just saying, because well, actually that milling machine wasn't working and so that data is not valid. We've got to go clean that up.

Marco

⁓ I think the lack is the context of the data we have in manufacturing. So we have lot of data, but mostly the context is missing and then it's hard ⁓ to understand some data fractures. Why the data isn't in that way. And the second thing is, think the users within manufacturing, the CAM users for example, they have so many things to do and at one time they have to do the machine setup. Maybe they have to make the first part. And then when you have a system with a lot of questions during your work, ⁓ they are not happy to give a lot of feedback during their work. so ⁓ I think ⁓ when we have more context data, then it's easier also to use AI to clean up data. ⁓ But for that, we need definitely the context.

Michael Finocchiaro

Very good.

Tanmay Aggarwal

Michael, I would want to just go back to that question of probabilistic as a deterministic in a different perspective here, which is I really like what Mark was saying. If you think about the process planning stage of the use case, there, for example, one way to handle deterministic versus probabilistic is by giving choices to the user.

Michael Finocchiaro

Absolutely.

Tanmay Aggarwal

And that allows one approach to go around that. And that's kind of what we do at Lambda, right? We, rather than saying, hey, here's the one and only way you can go about machining this particular feature, here are the four different ways you could go about doing it, but the last mile context we're missing. So we're gonna let the user kind of decide that. So we're gonna go from that. AI kind of giving you a probabilistic approach and then the user kind of turning that probabilistic output into a deterministic selection. So that's, think, one thing that I think works well. The other is when you think about, at least in the world of Lambda, we connect to the shop flow downstream and we get the data out of CNC machines. A big reason of doing that is because there's inherent process variability in CNC and manufacturing.

Michael Finocchiaro

Okay

Tanmay Aggarwal

Because there's inherent process variability, that by nature, by design, is probabilistic in nature. So that turns into a very good use case to apply AI. you are, I mean, actually the current domain of manufacturing struggles with the fact that you take a probabilistic environment and try to force-fade it into deterministic.

Michael Finocchiaro

Right.

Tanmay Aggarwal

system and that's where the problem often is because there's a probability that a tool may break. There's a probability that this part may go out of spec given the current condition of the machining system, but we don't have the data or the visibility into that nature of that process. And we kind of have to live with the deterministic outcome that that creates at the end. So I find that to be an area where the probabilistic aspect of AI actually is a very good strength in that scenario.

Michael Finocchiaro

That's interesting. We'll probably see more research around that collision of those two worlds and where those things are actually with the public. That's a really good point. The fact that that probabilistic thing is something to take into account and not just discard for those reasons. That's cool. So like when I was also wondering, are you guys using LLMs that you've trained via rag or do you have the idea of I don't know if you saw the podcast I did with the Mauer Farid of Leo AI, where he actually created a large mechanical model. He created his own by adding all this mechanical engineering knowledge to a large language model that he trained. ⁓ How are you guys doing? How are you adding, are you just using RAG to add, you know, knowledge sets of information you've gotten from all your experience or are you training a specific LLM around your data set?

Marco

I think when it comes to LLM, ⁓ we are using standard models. What we did in the past was, especially regarding the geometry recognition, that we created synthetic data around geometry and built up our own models, but that it's not really LLM, that was more for feature recognition. ⁓ But in general for LLM we are using more the standard models that are in place in the market.

Tanmay Aggarwal

Yeah, I mean, in our case, again, historically and even now, like a lot of what we've done has been, you know, traditional ML, like, like decision tree models is something we use very, very commonly downstream, we use things like LSTM models and other kinds of ⁓ techniques. We are, though, very explicitly looking into the rag approach, Michael, for introducing ⁓ LLMs and LRMs into the workflow. ⁓ But that's kind of something where we are in the early days of looking into how to incorporate that relative to the other techniques we're already using.

Michael Finocchiaro

Because I imagine that would be also a big change to your tool chain because you'd have to have LLM Ops and you'd have to change the DevOps chain to include all that stuff, which would make it little more complicated, right?

Tanmay Aggarwal

Well, thankfully for us, least I think from day one, we built it in a very modular manner. it's actually everything that we've done fits very well into the whole agentic AI kind of framework and architecture. So we see kind of that being the evolution for us, where whatever we have so far are our individual components that fit within a broader orchestration there.

Michael Finocchiaro

Like a MLOps is already there. So we've covered like how you guys perceive AI, how you use, your developers using it, how it's used in your products. Now as we head towards the end of 2026, and we just saw today Jeff Bezos announced Prometheus, industrial metaverse company, 6.2 billion investment, which was kind of interesting because I'm doing a series of posts about the industrial metaverse. Where do you guys stand on AI today? Looking at ⁓ the end of 2025 into 2026, ⁓ are you still super bullish? Are you kind of concerned? Where do you think it's gonna go? What are your thoughts about it? Neither of you can pick that up, I don't mind.

Tanmay Aggarwal

Narka?

Marco

Yeah, I think in general when it comes to AI in general I'm quite bullish. I think that's a ⁓ very interesting technology and we will see a lot of changes in the future based on AI. ⁓ Regarding manufacturing, I'm still skeptical because I'm not sure if the users or the companies are ready for AI. ⁓ because I think they are searching for AI because everybody is talking about AI and it's important to use any kind of AI. ⁓ But maybe that's not the correct time. ⁓ And maybe it takes two years later to start with AI and manufacturing. ⁓ That comes to the discussion we had earlier regarding deterministic and probability. ⁓ And I think that at the current state in manufacturing, the people are feeling better. to have not this black box approach where they put something in and have something out and don't understand. think the technicians, they like to understand and they like to sit in the driver's seat. ⁓ And so I think for real AI applications, as we see it in other industries, ⁓ it will take some time in manufacturing that the users are... and they would like to understand it. ⁓ So I think in that industry I'm more skeptical, ⁓ especially in Europe and Germany. ⁓ In a broader view, I'm definitely bullish in AI.

Michael Finocchiaro

Well, maybe if you don't mind me kind of ⁓ drilling down on that, like, do you think that there's a difference then from smaller manufacturers to larger ones? Would the larger ones be maybe more aggressive? They've got more money to burn, whereas the smaller ones, you if they'd make a mistake, they could be out of business. Is there a bit of a difference there or not really?

Marco

No, I think maybe it's the way around. ⁓ We see some smaller companies and maybe when you have one person in place at the moment who is responsible for CAM and for the quoting and for ⁓ the first part on the machine, for him it's easier to understand the whole process. In bigger corporates where you have different departments, you have a lot of ⁓ talks between the departments and

Michael Finocchiaro

Okay.

Marco

I think for them it's maybe harder in the context of manufacturing. So I see here SMEs more flexible.

Michael Finocchiaro

Interesting. You have that same experience. mean, Marco and I are sitting Europe. You're sitting in the States, Tenway. Probably you have a different perspective.

Tanmay Aggarwal

Yeah, it's interesting. I think you'll hear a different perspective. ⁓ For sure. I see that kind of definitely, of course, continue to be very, very bullish. I think we are in the early, early days. We have not even scratched the surface on what's to come. ⁓ I fully agree with Marco in the pragmatic nature of, of manufacturing as a domain and, and thereby I think like any, any new technology, there's going to be an adoption curve, right? And that's gonna, you know, I look at it like the way maybe robotics was in manufacturing 15 years ago. It's kind like maybe where AI is today in manufacturing. like, you know, it's taken 15 years for robotics to become like mainstream. I think at this point, hopefully I can say that without offending anybody, but I think, I think it's time to become more mainstream now.

Michael Finocchiaro

Hmm.

Tanmay Aggarwal

I hope it's not going to take 15 more years for AI to become mainstream in many fashion, but I do see a very clear ⁓ demand pull that's emerging in market. Customers are definitely in the awareness and exploration phase right now, ⁓ but you can start seeing the early adopters and innovators reaching out, looking for technologies, which was not the case three years ago. I think... ⁓ I think thanks to the whole chat GPT movement and whatnot, ⁓ it has created this sense of awareness, which is causing the early adopters to get out of their comfort zone and look for technologies like probably what up to pass provides or Lambda function provides. And we are starting to see that demand pool and we are seeing that at a global level, not from specific geographies. It does tend to concentrate in certain segments of the market. I think the behavior patterns of an OEM and a large tier one and a tier two is very different than a behavior pattern of an SMA. And to your point, Michael, I do think that the budget allocation and dollars available to explore and adopt some of these technologies are going to vary quite meaningfully. And I think then it comes down to like your own playbook. you know, who are you focused on your approach to bringing a new technology to your customer base? But I do think there's demand probably in both ends of the spectrum, but probably very different demand. So like, it's very, very clear that I think that you cannot go in with the same playbook, try and serve the SMB category as well as the large enterprise category, because I think the needs are very, different.

Michael Finocchiaro

All right. That makes me want to ask, obviously the market is dominated by three players, right? The big three. ⁓ And yet, in my studies, I've found 380 plus startups from CAE to CAM to ⁓ Kishore Threaded, MBSC, ⁓ new kinds of doing design. I mean, it's insane. And you would think, well, the big three should be doing all that stuff. So there seems to be a gap. Is that because they just are too slow to catch up and you guys are much more agile and you can get the solutions out much faster with more later technology and you're taking advantage of that? I I'm just, I'm not sure that two years ago there were 380 companies doing startups in engineering software and industrial AI. I don't think there were more than 200 in. I'm probably only scratching the surface. There might be 450 out there. There might be another 80 I haven't found yet. I every day I found another five or eight or 10. It's just an interesting dynamic. How do you guys see that? you guys see, you guys, does that excite you guys? say, well, I've really got this opportunity and I want to grow and become the next Siemens, Comos or the next Semitech. Or you're thinking, no, I want to be acquired by one of those guys. You know, is it, how do you guys see that? Do you think that the, in other words, are we moving towards? These monoliths will end up acquiring you guys and become remain monoliths. Are we starting to chop away at the foundations of this monolithic system and go to something more agentic where there's a lot of room for everybody to play and get a best of breed solution for whatever I'm doing, whether it's manufacturing or design or ideation. Sorry, kind of big question, but I just wanted to see how you guys would react.

Tanmay Aggarwal

Yeah, maybe I can kick off Marco and I'm curious to hear your perspective on it. But first of all, I'm surprised it's only 380. I want to know a lot more. think but that's great. And it's amazing that you're starting to map it out. think that's great. So but I think think the large incumbents, at least from our lens, the way I see it and the way we've been seeing it over the last few years, I think they've taken they are taking a much more platform approach.

Michael Finocchiaro

Yeah, that's probably more.

Tanmay Aggarwal

It's a broad domain, right? It's a very, very broad domain. And so if you're at the size of one of these large companies, they are looking at a much broader problem statement. They're looking at tapping into much bigger budget pools of the customer profiles. Well, startups like us are trying to identify specific use cases in that broad field where we can come in and provide, you know, hopefully, a meaningful amount of value creation for the customer to create our own market space. then, and I think that's what's creating a bit of a synergistic relationship between companies like us and the larger players where, again, we're very grateful today, like, you know, we're partnered with Siemens and Autodesk and Mastercam and others, and I think we couldn't exist without those relationships. So again, very grateful for these large players to allow startups to innovate and play. And I think that's the synergy as I see it, where we can continue to be more focused be the more, to your point of the agenda topic, we could be the more AI agents that are designed to solve specific use cases while maybe those companies take a much more of the orchestration layer or the more foundational model problem statements which are outside the realm of any individual startup, maybe at the stage we're at. So anyways, Marco.

Marco

It's definitely a tough question and I think that we will see a consolidation of all these startups in the future because of two reasons. One reason is the complexity in technology. So for example, as we started, we also tried to make for example the Toolpath generation on our own by using some SDKs from well-known vendors and like to create a new CAM system within the cloud. and last year we decided that it's not possible to develop 30 or 40 years of experience of existing CAM infrastructure like MasterCAM or GibbsCAM or HyperMIL to do a new approach of CAM programming within two or three years, because it's too complex from a technology standpoint. And the second thing is... nobody is financing this development to try to put all this experience into the cloud. mean, we see with Onshape and PTC, we see that one of the big corporates in the CAD design, they are doing something into the cloud because they have experience in their legacy field and then they transfer it into the cloud. And I think that's an ⁓ approach, what is possible. ⁓ for up to parts, I see our benefit that we had from the beginning a holistic approach where we are like in the middle layer between different CAM and CAT systems. So we are not ⁓ fixed to one of the systems and it's like you have the possibility to create your company individual knowledge into a rule-based AI system. And when we, for example, when we are creating CAM projects, we are not creating CAM projects for Mastercam or for GibbsCam or for HyperMill. So we creating a CAM project based on your corporate knowledge. And then you are able to decide ⁓ if you'd to transfer it to Mastercam or to GibbsCam or to HyperMill. And I think that is our approach. That's what will help.

Tanmay Aggarwal

you

Marco

us to be part of this ecosystem also in the future. But in total I see definitely a consolidation in the market because it's too big and as you said big companies like Siemens, Sandvik and so on, they are all working on that stuff and for sure they will find technologies and products within their own applications. And then maybe some of the really exciting startups are being bought or being replaced or from developments out of the big companies. But I think that's a common evolution in every new innovation field or a branch.

Michael Finocchiaro

Well, yeah, because we've seen over what the last couple of years, huge consolidation and simulation, right, with Altair ⁓ and MSC going ⁓ from Hexagon to... ⁓ it went to Synopsys. No, was to Cadence, Cadence, the MSC, the sale of MSC from Hexagon. And so, I think that I agree that there... some consolidation might be happening in the cam field next because obviously we've already done the CAD stuff all rolling up to the three companies so far. It's pretty interesting. So if we switch gears to the customers, which we already started addressing to a degree, when you go to your customers, and if you look at digital maturity and a spectrum of one to five, one being we still use Excel and we're still using email to... I've got this fully agentic adaptive digital twin that's, you know, be able to repair itself, which we're not at yet. won't maybe not be for a long time. ⁓ I would call that five. So I'd say, you know, the general customers are somewhere between one and three. I mean, maybe there's one or two in the world might be over four. Is that the same experience you have? mean, how do you, are you able to assess that? Because, you know, it's It's an interesting thing because that's of course where you guys are bringing value when you bring in an AI powered solution. It's going to hopefully help them mature. So who wants to pick that one first?

Marco

I can go for it. So I think in average ⁓ we are on a two when we see ⁓ the customers. ⁓ I think the most important tool in all manufacturing companies is Excel. ⁓ Excel is everywhere ⁓ from small SMEs ⁓ to big corporates. ⁓ And that was something we realized ⁓ in the beginning as we searched for structured data.

Michael Finocchiaro

Okay.

Marco

by our customers and we saw that today the structured and especially context data is not there. And so we positioned Afterparts now also as a possibility to set up structured and context data in manufacturing companies. ⁓ Not focus on AI from the beginning, but have the possibility to store all your data at one single point of data. ⁓ And so I think then we have the possibility to help our customers to increase their ⁓ level in this question you asked. But I think in average, yeah, by a two, ⁓ for sure we have also some prospects, some customers, they are on a three or on a four, but that's the minority.

Michael Finocchiaro

How about 10 million in the States, do you have the same perspective?

Tanmay Aggarwal

Yeah, would say very similar, Michael. I think it's anywhere between a one and a two. And I was looking at it from the lens of you have, I think customers have made a lot of progress on traditional automation. But when it comes to embedding intelligence into the workflow, those are two different axes, if you may. of a digital maturity framework, right? So there's a lot of progress been made on automation, right? You see a lot more robotic cells. You see a lot more of go-bots are becoming more common, especially in the more advanced shops and more at scale shops. So there, they are maybe at a three on an automation scale. But when it comes to embedding intelligence that can unlock autonomy in their workflows, they're maybe still close to a one or a two. And then you can of pair it back from there. Like I would say that's by the most advanced. We'll look backwards from now.

Michael Finocchiaro

But I, okay, I like that aspect of a lot. When I think of it, I'm thinking more is like around the data silos, right? The fact that, you you've got the PLM over here and the RP over here and all this cam data sitting on somebody's desk on a bunch of printouts. ⁓

Tanmay Aggarwal

We are at negative five, Michael. We're probably at a negative five in that world. ⁓

Michael Finocchiaro

What's that? Yeah. Well, it's funny. You would have thought you would think that Formula One made me being as awesome and how much money they have would be like a five. And yet when I had Michael Rosen from Quixon last week, he was like, no, no, no, you don't understand. They just had this data like basically everywhere. The test data is just, which is just insane, right? Where in 2025, you would have thought we would have gone further.

Tanmay Aggarwal

Yeah. I would say, yeah, yeah. But I mean, not surprising, but I mean, I would say like from our experience, like we tend to work very closely with enterprise customers, right? Like that's where we focus on. So you're talking about some of the largest organizations, leaders in their own industrial supply chains, and they are the ones who are starting to talk about how can I bridge my data across these different silos, right?

Michael Finocchiaro

Wow.

Tanmay Aggarwal

And there are a lot of just fundamental reasons why it's hard. yeah, I mean, again, these are heterogeneous environments, right? No company has a homogenous suite from PTC or Siemens, right? The reality of the matter is within a company and across their supply chain where the data needs to be exchanged, you're going to have a mix of different systems.

Michael Finocchiaro

Well, the semantics are different, the transaction versus collaborative.

Tanmay Aggarwal

⁓ So data connectivity, think even at the most advanced companies is still at a pretty low level of maturity today, but you are starting to see very clear motivation and dollars being spent to unlock that connectivity layer. which I think is the foundation to doing anything more intelligent with that data. So that's kind of where we are. So then if you think about all the other companies in the supply chain who are not as big or as advanced, I mean, they're really in their own data silos. And I think for the next 10 years, it's probably going to be this, you know, this could be a slow journey towards more integration, which would be nice for, I think, companies like us because that creates opportunity.

Michael Finocchiaro

Yeah. Yeah, obviously. Well, I was also thinking that the most mature ones would have a chief data officer, which is not the same as the IT guy, right? Someone just managed the data with a matrix of data owners and all the different BUs and then data stewards and data custodians. Just the kind of organizational changes you need in order to make that happen. Because if that's all under IT, then it's just an IT thing and it's just another solution rollout. And when you talk about data, it's not a solution. It's a whole mindset. It's a It's a revolution. Sorry, we didn't let you answer that one Marco. You want to jump in?

Marco

⁓ No, ⁓ everything good. I just have an additional thought also that you have when it comes to AI manufacturing, there is not one AI in manufacturing and that's also the difference between Lambda function and its other parts. I the approach from Lambda by having the feedback from the machines and have a look into the tool selection and to the toolpath and then use AI to optimize that, I think that's That's one topic of AI. Then ⁓ you have a topic of AI in the design phase, where you can maybe optimize your design by doing it more cost-efficiency. And then you have ⁓ our work preparation workflow, where you have another scope on AI. And so I think there are also different fields where you have a different level of AI readiness. ⁓ Or it's always also important ⁓ what you like to achieve with AI. against what are you optimizing. ⁓ And I think in Lambda Functions field it's maybe today already easier for the customer to define a scope what you like to optimize in that field. And in Repreporation it's maybe harder because you have a lot of different data. As a Chop Shopper you have no own design department, you have to rely on the designs of external companies with a bunch of different designers and there is more complex ⁓ to have a general standard of context data and that's also a thing we have to keep in mind that when it comes to AI in manufacturing there is no single solution or way to work with AI. data silo or every department has to define his own way of working with AI. That was just an additional thought I had ⁓ during the talks.

Tanmay Aggarwal

⁓ I couldn't agree more, ⁓ Marco, 100 % agree. I think, Michael, like where I'm having a lot of conversations with customers about is, particularly, again, enterprise customers and we're dealing with C-suite executives, is if any organization is truly serious about adopting AI, particularly in the world of industrial and manufacturing, the number one question to ask is, have you thought about what budget you intend to allocate towards this? And 99 % of the time, the answer is they don't And that's quite indicative, I think, of the current state of AI maturity in this domain. People are starting to, they want to learn about it. They want to do something about it. I think they're very clear at a strategic level. They understand that they need to adopt.

Michael Finocchiaro

Hmm. Hmm.

Tanmay Aggarwal

these new technologies to avoid getting left behind and to stay competitive and to stay being a leader. But I don't think that entire shift in the market is not there. There's a very clear defined budget on the P &L, which is clear where that money is going to come from, who's going to fund it. ⁓ I think they're in that exploration phase. think most companies are still trying to figure out. And then that's okay. But I think that's that evolution over the next 12 to 24 to 36 months. I think as that that clarity on use cases, that clarity on ROI comes in when people start experimenting with with low cost tools, I think we'll start creating a more of a more structured marketplace for.

Marco

Yeah,

Michael Finocchiaro

But.

Marco

and that is something you mentioned earlier Tanmay and that's definitely changed over the last three years and we recognize it also in the Emo that three years ago the people are coming to the booth and just reading after parts, AI assisted work preparation, what's that and what can you do for me? And on the last Emo now in September, the people come to the booth with, I know you, I know your competitors, that's my use case, that's what I would like to achieve. I think that's the reason why they don't have a budget, because they have no goal at the moment defined. And that's still in the exploration phase, because when they say, we would like to use AI to decrease our tool costs by 10%, then they can go into their costs and can do a calculation what they're saving in money. And then they can say, okay, then we have this budget. But today they are still, as you mentioned, in this ⁓ exploration phase, but it's improved the last three years from

Michael Finocchiaro

Mmm.

Marco

what's AI manufacturing into. I heard from AI manufacturing and it's interesting and I think the next three years we will see that's our use case, that's our goal, that's our budget.

Tanmay Aggarwal

Exactly.

Michael Finocchiaro

And do you think that implementing Lambda function or up to parts is one of those accelerators to getting to digital maturity? When you put in a solution as the customer, it get a bit of an epiphany? Does it push them to become a bit more mature in their data in order to get more out of the tool and get to those KPIs faster than they would have otherwise?

Tanmay Aggarwal

I can start here. I definitely think so, Michael. There's definitely an element of, I didn't realize I could save this money that quickly. And so for us, a lot of times, the focus is, what's the quickest ROI? Where can I show time to value? Reduce the fastest thing I can deliver value. And 99 % of time, it may not even be with AI.

Michael Finocchiaro

Right.

Tanmay Aggarwal

Like, you know, it could, maybe that value is just in looking at data a different way. Like it's just basic stuff, but that starts opening the mind of the executives and the mid management and everybody else to say, okay, there's more here than meets the eye. I think, then I think we see it as like, it's a stepping stone. Like I don't think Lambda function is the one all be all solution at all. Like it's one tool in a broader toolkit.

Michael Finocchiaro

and pay something else. One piece.

Tanmay Aggarwal

To Marco's point, and I mean, I think there are many different elements that will come together. But I think we do see that the tools like Lambda and hopefully and up to parts and others, like just opening the market for a broader adoption wave. But I do think it'll still take time. is not something that suddenly, AI is not going to take over the world of manufacturing tomorrow. But I think there is lot of impact and potential to make life easier for people. But Marco, please.

Marco

I see it the same way, we definitely have prospects for customers. They have a real benefit from the beginning, because ⁓ maybe they have a niche or anything else where it already fits very well or they have the data for that. ⁓ But I think the same what Tenmei mentioned is that sometimes it's also helpful for the prospects ⁓ to fail with an AI project, because then they see what they are... what they are lacking. So they heard about AI, they would like to do something about AI and then they do an evaluation call and then you ask them, do you have a tool management system? And tool management system or tool data is mandatory for a solution like us. And then they say, no, we ⁓ buy the tools as it goes. So what's on the market? And then you can say, okay, then our solution will not help you because we need proper tool data. ⁓

Michael Finocchiaro

All right.

Marco

and you need proper tool data to evaluate our solution if we select the correct tools and choose the correct tool data. And then that's also a helpful step to increase the digital maturity ⁓ to see, okay, I would like to do something with AI, but I have to realize that I have to do some groundwork. That's the same with robots on machines.

Michael Finocchiaro

some homework.

Marco

So just put a robot in front of a milling machine, that doesn't mean that you are ⁓ doing 24x7. You need standards. Automization always needs standards. And ⁓ I think that's also something we and all the other AI companies in manufacturing ⁓ can give like a mirror to the prospect, to the customer and say, okay, yeah.

Michael Finocchiaro

Mm-hmm.

Marco

It's good that you are ⁓ thinking about AI, but you have to do some homework.

Michael Finocchiaro

Wow, I love these answers. These are really great. Thanks guys. Just before we go, we got a couple of minutes and I wanted to, because I do get younger people on joining the podcast and listening. And I know there's a lot of AI anxiety, know, like, you know, the AIS can take my job. Why am I even bother to study? Because what I'm doing is to be done by an AI agent anyway. ⁓ What kind of advice do you have to these younger engineers that are just starting their career? You know, all three of us have a long. careers in this industry? What kind of advice do you have so they stay relevant so that they, ⁓ you know, that there's going to be able to contribute in the years to come?

Marco

Like I start 10 way.

Tanmay Aggarwal

Go for it. yeah, no, I can't. Yeah, I have to. I would stop by saying, definitely no advice. I'm no one to give advice. But what I can share is, as an engineer, hopefully our job is not to execute tasks. Tasks are what AI agents are designed to do.

Marco

Okay, you can start.

Michael Finocchiaro

⁓ He's still thinking about his answer.

Marco

haha

Tanmay Aggarwal

are, you know, the question is identifying the right problems to work on, framing the right problems to kind of think about solving, and then building AI agents to solve those problems for you. So I think if you just reframe that perspective, ⁓ I think it makes it for a very exciting time as opposed to a scary time. I think, but yes, change is always going to be different. So we got to keep an open mind. I'm ⁓ a... big fan of constantly trying to just learn and see what's happening next. And I think there's a huge opportunity for the engineers who coming out now to do 100x more output in a much shorter amount of time, just because of the tools that are available to us. So yeah, I would just look at it from the lens of. our job should be to let's try and find the problems that are worth solving. And then let's think about creative ways to approach those problems and then use AI to the hilt to help create leverage in solving those problems faster. So that'd be my.

Michael Finocchiaro

Thanks, Tim.

Marco

I think the engineers, all the people in manufacturing, ⁓ they have it in their own hands. ⁓ You can use AI tools ⁓ just to rely totally on them and use AI as a singing point of truth. Or you can be the leader or sit in the driver's seat and use AI as an additional tool. And maybe that's also what Tanmay mentioned or meant to... to use AI tools to make more workload, to improve your work results, to use AI also as a learning platform ⁓ to increase your experience and so on. ⁓ I think when you always ⁓ doubt the status quo of the answers of the AI, ⁓ that's maybe a... not advice or ⁓ just my opinion ⁓ that we are humans. should not only rely on AI and take AI as the real truth and always be reconsidered the answers of AI and phrase my own knowledge, assisted, supported by AI, but not give everything into AI, because when I give everything into AI, there's no wonder when I'm being replaced.

Michael Finocchiaro

Yeah, it's like the so easy to just let it when it says, would you like me to do this? And you just say, yes, it's very, it's very easy to delegate that all the time. Oh, I've had a great time in this conversation. I think I've learned a lot. I appreciate you guys joining me. Tanmay and Marco, it's been great. For those that are listening live, there's another podcast on Friday with No Space and Infinite Form. This can be great. And next week. I'll have TD Engine and Ops Mate, which AI, which is going to be exciting. ⁓ So anyway, I wanted to once again thank Marco and Tanmay so much for joining me. And we'll see you guys next time. So thank you.

Marco

Thank you, always a pleasure.

Tanmay Aggarwal

Thank you for organizing.

Marco

Yeah, yeah. Take care, see you soon.

Michael Finocchiaro

I'm just going to. okay.