Building Physical AI for the Industrial World

Harj Taggar 0:00 [upbeat music] Today I'm excited to be joined here by Igor Marinelli, co-founder and CEO of Tractian. Uh, Tractian builds physical AI for asset-heavy industries and works with some of the biggest enterprises in the world, like Eli Lilly and McKesson. Uh, today they're announcing their, uh, Series D, uh, which values them at $1.4 billion, and they've crossed over $100 million in ARR. So congratulations, Igor.

Igor Marinelli 0:29 Thank you.

Harj Taggar 0:30 Cool. Well, why don't we just start, like give us a sense of, um, what exactly does Tractian do? What's the product? Um, how does it work? Just tell us a little bit about that.

Igor Marinelli 0:38 Yeah, for sure. We're tackling essentially the trillion-dollar problem, which is machine downtime. So think about a data center, right? Um, everybody's talking about GPUs right now and-

Harj Taggar 0:49 Yeah

Igor Marinelli 0:49 ... and capacity, but have you ever seen a data center, you know, I guess from, you know, from a distance where, where you can see all this wrapper, right? Which is, you know, from the moment that the energy's coming in, it's, it's, it's coming to a transformer and then getting distributed to, you know, battery packs or mega packs and then going through chillers and then evaporators and diesel generators. So we are essentially the key monitoring infrastructure for critical assets and, and we do hardware and software solutions and AI to tackle those. Um-

Harj Taggar 1:27 By the way, if I'm building software, I have observability tools, so I know kind of like if they're, like in the logs if something's crashing or if someone's visiting my website. You guys are, are that for the physical world.

Igor Marinelli 1:36 Exactly. So, you know, within the largest, 500 largest industrial operations today, it's, it's sort of estimated that 11% of their revenue is blocked.

Harj Taggar 1:47 Oh.

Igor Marinelli 1:47 So it's the cost of opportunity, which is $1.4 trillion in, in, in, in industrial GDP that is essentially blocked because assets like chillers, HVACs, you know, think about what's cooling this room right now. It's probably in the, in, in somewhere in the ceiling here. Uh, this could also be monitored by Tractian. And so what we do essentially, we guarantee 100% capacity and availability for that physical asset.

Harj Taggar 2:13 And so, and what exactly is the product? Like is it, is it hardware? Is it software? Like just help us understand kind of what exactly-

Igor Marinelli 2:20 So, so I like to think like, you know, y- you've mentioned two of our customers, right? So think about Mounjaro for a second, which is this miraculous drug, right? You know, the moment that Mounjaro is being manufactured, a- apart from Mounjaro's raw formula, which is sort of trade secret, it needs compressed air to be manufactured. Uh, it needs, you know, palletizer robotic arms, uh, to continue along the, the journey of the line. So you think about these pieces and the reason why they fail, it, it, it is essentially three main things. It is a, a, a mechanical reason. Um, so you know, you know, the robotic arm couldn't actually work properly and then it caused dow- downtime because, you know, uh, one of the joints were not, uh, uh, uh, lubricated correctly.

Harj Taggar 3:10 Mm-hmm.

Igor Marinelli 3:10 Um, so we've identified that. But it could be as well the second re- reason, which is electrical. So there was a discharge or a surcharge that essentially, uh, led to one of the electrical components to, uh, burn out. So we also identify that before it happens. Or it could be an operational reason, right? An operator went there and kicked the thing or, you know, kicked the box and, and now, you know, the palletizer is not, is not doing its thing properly. And we also have ways to identify these, right? So it starts with, you know, proprietary hardware essentially, which is cool, right? We have, um, anywhere from vibration, ultrasound, magnetic field, current sensors that are beautifully designed and non-invasive but, you know, they're there for the purpose of collecting that data and then that data gets transferred to, um, our AI which, you know, it, it, it has a bunch of, uh, deterministic properties like, like Shazam, right? So you're, you're-- w- we are, you know, decompounding this analog data and, and assessing essentially, um, what is the, uh, actual condition that is happening with this asset. Just like Shazam is trying to assess if you're hearing Metallica or Guns N' Roses, right? And then later on, other models are going to try, try to assess based on the music that's playing, what is the volume? Is it critical? Should I go there right now? Can it wait? And that is the moment there that we, you know, make the use of course of like post-trained LLMs because then you need context. You need the manual data sheet of that asset. You need the PDF from-

Harj Taggar 4:59 Yeah

Igor Marinelli 4:59 ... the OEM, the original, uh, equipment manufacturer. You need a layer of memory because what's critical for you may not be critical for another type of factory or another type of user.

Harj Taggar 5:10 And presumably every factory is probably g- it's almost like, like it's like its own system, so it's probably got its own history and environment and, um, like context that's specific to that like factory or plant-

Igor Marinelli 5:20 Yes

Harj Taggar 5:20 ... that, that is useful to know.

Igor Marinelli 5:21 So, so it's, it's just, you know, it's very, very, very sophisticated but i- in its nature it's doing a very simple thing which is providing information for our key users. So information ahead of time and you need all of this stuff in order to make

Igor Marinelli 5:38 that happen.

Harj Taggar 5:38 Cool. Well, you're doing it, um, at like significant scale now. Like these are some of like the probably like the biggest like just plants and physical locations in physical asset, you know, um, destinations in the world and you're doing it I think like at more scale than anybody else at the moment. But if we rewind a little bit-

Igor Marinelli 5:55 Mm-hmm

Harj Taggar 5:56 ... um, when you went through YC in 2021, um, what was the original idea and what did you remember, what kind of what you had then? Like wh- where did it all begin?

Igor Marinelli 6:05 Yeah. So first we've applied, and I think we were-

Harj Taggar 6:08 Oh, no. [laughs]

Igor Marinelli 6:09 ... uh, we were zero in revenue, right?

Harj Taggar 6:11 [laughs]

Igor Marinelli 6:12 So, um, uh, for those that don't know, um, uh, I believe in winter of '21 was our second application.

Harj Taggar 6:20 Okay. Okay.

Igor Marinelli 6:21 So-

Harj Taggar 6:21 It's the first one I saw, so-

Igor Marinelli 6:22 Yeah.

Harj Taggar 6:22 [laughs]

Igor Marinelli 6:23 In, in reality is the first application, you know, in order to, you know, sum up all those parts, it, it really, like, took a lot of time for us just in order to test product market fit. You... I remember you mentioned to me once, and it, and it's, it's real up until now, like, Igor, you know, you, you're not gonna be able to deal with an MVP, minimum viable product. It has to be a minimum awesome product.

Harj Taggar 6:47 [laughs] Yeah.

Igor Marinelli 6:48 Because, you know, in order to test that, you're going to have to build your own hardware, and then you're gonna have to build your own software, and then you're gonna have to build your own AI. So it sort of had, had always had, like, a high ceiling-

Harj Taggar 7:00 Yeah

Igor Marinelli 7:00 ... in order to cross that. And-

Harj Taggar 7:02 'Cause, 'cause it's like, it's real machinery, it's real stuff, and it's expensive. And, yeah, I remember this conversation. You can't have something like, "Oh, like, we missed it," um, and ever expect to get, like, the, the trust back with that customer.

Igor Marinelli 7:13 Yeah. Yeah. And with that, it, it, it truly came, uh, essentially a little bit of the chicken and the egg, uh, process in the beginning, of sales especially, because in order for the system to work, it has to have a lot of data. In order to have a lot of data, you know, everything has to be ready, and in order for everything to be ready, a customer needs to trust us, that like, "Okay, I'm going to give you my 100 most precious piece of equipment-

Harj Taggar 7:37 Yeah

Igor Marinelli 7:37 ... and it's now ran by Tractian." Right? So, so that whole process, [laughs] you know, it gave us a lot of gray hair, for sure. Uh, and YC truly was the first believer, but, you know, in order for that to be accomplished, it, it, it, it has taken us, like, two years from, uh, since foundation-

Harj Taggar 7:57 Mm-hmm

Igor Marinelli 7:57 ... in order to have, uh, your, your yes, right? And, and, and so we were moving every piece that we possibly c- possibly could to have product as much ready as possible. And, and that was generally done by founders because we were all technical founders, when in fact my first hire was actually an SDR, so sales development, because we thought, "Well, we can do all this stuff. We're computer engineers. We're gonna figure it out, but we need someone to sell." Because, you know, in our industry, you know, one thing has to, has to run, you know, in parallel with the other, right? So, um, and, and there we were also generating pipeline from, from day zero.

Harj Taggar 8:40 Cool. So usually with this type of idea where, um, the sort of bar to entry is high because the product needs to be quite good on day one, um, uh, and it can take, you know, a coup- a couple of years to really perfect it, um, founders just need, like, a very high conviction in what they're doing. Um, uh, you need, certainly need higher conviction than something you can sort of, like, build over a weekend and rapidly iterate on. So maybe tell us a little bit about how did you come up with this idea? So what was the, the founding story of the idea, and, and what gave you such conviction and attachment to the, to

Harj Taggar 9:13 solving it?

Igor Marinelli 9:13 Yeah. It, it was a combination of two things. The first one is that I've seen my father struggling with that.

Harj Taggar 9:22 Yeah. Tell us more.

Igor Marinelli 9:23 Throughout-

Harj Taggar 9:23 I think it's a fascinating story

Igor Marinelli 9:24 ... you know, throughout this whole career, he was a plant manager of a, a paper manufacturing facility. And, you know, not only would I not see, you know, see him very often, um, he would also complain a lot about, you know, the technologies that we, he was, uh, uh, was using at that time. And at the same time, my co-founder's father, sort of like he lived the same thing, where he worked, like, 35 years in this cement factory. And so, you know, the first thing that's fascinating me-

Harj Taggar 9:55 My parents too, by the way.

Igor Marinelli 9:56 Really?

Harj Taggar 9:56 Yeah. My parents both emigrated from India to England. I think I've talked about this, but yeah, they both, like, the only jobs that were available for immigrants in that era were really, like, machine operators in factories. So my dad worked in the same paint factory for almost all his, like, 40 years-

Igor Marinelli 10:10 Wow

Harj Taggar 10:10 ... something like that. And my mom in, like, a plastics factory putting together components. So yeah, it definitely-

Igor Marinelli 10:15 So you know.

Harj Taggar 10:15 Yeah, yeah.

Igor Marinelli 10:16 Yeah. So, you know, it, it is like, my fam- family came from Italy, and y- y- essentially manufacturing was, right? And, and it still is how you build, you know, this, this very, very, you know, structured middle class, essentially-

Harj Taggar 10:30 Mm-hmm

Igor Marinelli 10:30 ... which was, um, uh, my, my upbringing essentially. But that was it, right? Like, you know, sort of, you know, this s- Where sort of, like, I, I grew up and, and seeing his struggles, and then when we got together, Tractian has a very different story than other companies in the sense that in the day that we started the company, which was September 2019, I remember I bought 40 domains online.

Harj Taggar 10:58 [laughs]

Igor Marinelli 10:58 So, like, tractian.com, .com.mx, .de, .jp. I think I just didn't buy China.

Harj Taggar 11:04 And who is we? Sorry, it's funny.

Igor Marinelli 11:06 Yeah.

Harj Taggar 11:06 But who, who is we at this

Harj Taggar 11:07 point, right? Yeah.

Igor Marinelli 11:07 It was, it was me and Gr- Gabriel, essentially.

Harj Taggar 11:09 Gabriel.

Igor Marinelli 11:09 So that's the thing. We had a problem idea where we've seen that, um, you know, machine downtime was a big problem. We couldn't really, you know, you know, research a lot of attempts prior to, to what we were doing, right? Like, it, it's not that nobody has ever tried, but nobody has truly ever tried successfully. So it was this huge white space-

Harj Taggar 11:33 Mm-hmm

Igor Marinelli 11:33 ... and, and a problem that we were familiar with, and the butterflies on stomach in, in, in day one and sort of like, "Okay, from now on we're going to have to figure out what is the proper solution." Right? So I, I say it's different in that sense that it truly in, in day one, two or, or three, we were still thinking, "How are we going to solve this problem?" So I visited essentially 88 facilities in the first year. So it was just, like, on the road a lot, on the airplane a lot, talking with dozens of founders-

Harj Taggar 12:05 Customer discovery, basically.

Igor Marinelli 12:06 Exactly. And onsite is still a, a DNA of our, uh, sales culture up to nowadays, right? Just last year we've done, uh, 2,500 onsites. Uh, which are actually, you know, I've, I've, I've came to the conclusion after fighting in multiple board meetings that, look, um, it is actually cheaper than marketing, right?

Harj Taggar 12:28 [laughs]

Igor Marinelli 12:28 Like, you know, ta-take a look at the cost of, the growing cost of, like, inbound coming from Google, and, like, a, a, a, a Hertz, you know, r- car rental and flight and hotel. When you sum it all up, you can move that deal to a point that you've already surprised the customer, that you're there, that you're mapping their, their struggles, and, and we do this thing which is risk assessment, so we're going to do the job for you in mapping your whole facility and say, "Hey, here's the opportunity. Here's the assets that need-

Harj Taggar 12:57 Mm-hmm

Igor Marinelli 12:57 ... uh, um, observability, and this is the asset protection that you're going to need," and so on and so forth. So-

Harj Taggar 13:04 You basically do a bunch of work upfront, like, during the sales process for them.

Igor Marinelli 13:07 Exactly.

Harj Taggar 13:08 Cool. Okay. Well, then let's definitely come back to the sales process, 'cause I think you guys have interesting insights there. Um, on the customer discovery front, so you're going to these, you're going out to factories and, and manufacturing plants in person, and you, um, a- and how are you approaching that? Like, something I'm always curious about with founders is did you have in mind, "I'm gonna pitch them on this solution and see how they respond to it," or were you, um, just observing? And if, like, what, what was your process for figuring out, like, the, the right way to get use out of those, those visits?

Igor Marinelli 13:37 So I've always tried to validate if there are problems in different segments, you know, what are the sort of overlaps that I can find from a pharmaceutical line to a food and beverage facility to, uh, a, a, a paper, uh, uh, production facility. So when I was going on this, like, expedition, it, you know, the thought has always been, "I'm picking up all these problems," because you're showing up, right? And imagine even nowadays, showing up with, you know, sort of AI in, in, in, in a very, you know, you know, rigid, um, a- uh, environment, you know, you're, you're sort of showing up as a little bit of, like, you know, "This guy's going to help us to tackle all our problems," but it's, it's up to you to sort of, like, define the lane on up to how much are you going to solve. And we've always tried to avoid generalism in the solution and, and going as much as in-depth in sort of compounding, if we solve this problem for this person, is it the same domain of that other customer and that other customer and that other customer? That has always been our approach to a point that we were resisting, you know, adjacencies. So if you think about traction today, we're, we're very, like, domain specific in that realm, uh, which is weird because we're sector agnostic too, but just because we have the privilege that a HVAC, you know, it, it, it, it's the same equipment for domestic use that it is for industrial use, right? So it's, the, the segment is, uh, array is huge, but we've decided to focus on that specific thing.

Harj Taggar 15:25 You go in, you're learning about, sort of, you're sort of trying to map the problems between customers that you're seeing. Um, do you remember who your first customer was? Like, how, like, um, uh, how did you convince them to trust you, and, and what was the process of kind of getting up and running with them like?

Igor Marinelli 15:40 I do. I do. Um, they are, our customers, uh, up to nowadays-

Harj Taggar 15:45 Oh, wow. Okay, cool. Awesome

Igor Marinelli 15:46 ... they are this, you know, you know, hydraulic pump manufacturing facility. Uh, and I do remember because I've been there multiple times.

Harj Taggar 15:58 [laughs]

Igor Marinelli 15:59 Right? So, but what, what's funny is, like, you know, I remember, like, at that time I had, again, an SDR that I had already hired, a sales engineer that I had already hired, and, and, uh, an implementation engineer that was with me as well. So, like, how much are you guys investing today? Is it, is it, like, 500, 500,000?

Harj Taggar 16:18 500K per company. Yeah, yeah.

Igor Marinelli 16:19 You know, in my time-

Harj Taggar 16:20 Yeah, your time [laughs]

Igor Marinelli 16:21 ... it was, like, 125, right?

Harj Taggar 16:23 [laughs]

Igor Marinelli 16:24 But I've used every dollar in, in, in hiring.

Harj Taggar 16:27 Yeah.

Igor Marinelli 16:27 Like, every dollar went towards, you know, personnel. Um, because what I've noticed in enterprise sales is that, you know, some founders are doing founder selling for so long, but, but without any support truly, right? So, like, you enter their website, you, there's their Calendly there for you to book, and then you talk with them on the first stage, second stage, third stage. But, like, enterprise customers, they, they, they want to see that robustness from the get-go. They want to look at your LinkedIn and, and see that sort of, like, rise on employee count.

Harj Taggar 17:02 Yeah, interesting.

Igor Marinelli 17:03 All of that is kind of important so that you can open that first door, and, and, and I think, like, my best application as a founder, it, it, I, I consider myself a solutions engineer, right? So in order to generate demand, you know, it, it truly, like, there's no surprise, it's, it's scientific at this point. In order for you to generate leads, in order for you to, you know, have that first sales call, you're going to follow a script. There's, again, multiple methodologies too, but, like, you have to just come in, in, in crucial moments where, like, it, it is truly the v- verge of, like, sales and product, and there's that gap that you're filling, right? So, so I was doing that. It wasn't easy though, even with that sort of army of, of people, of, like, four people in that first-

Harj Taggar 17:50 That army, yeah. [laughs]

Igor Marinelli 17:51 [laughs] In that-

Harj Taggar 17:53 Did your first customer come through your SDR or did you find, find them?

Igor Marinelli 17:56 Oh, we, well, we hunt. We, we, we went there. We cold call.

Harj Taggar 17:59 Yeah.

Igor Marinelli 17:59 Right? Um, so that's, that was from day one, right? So we were, you know, and, and, and we are still calling, emailing, but it came from our, uh, our outbound, um-

Harj Taggar 18:10 Okay. And how did you convince them? How do you convince this hydraulic pump company to trust this four-person startup?

Igor Marinelli 18:15 Yeah. So there's, th- you know, throughout that whole process, which was long, it was tedious because- You know, they're trying to also gather social proof. We're talking about the first customer-

Harj Taggar 18:26 Yeah

Igor Marinelli 18:26 ... so you're not, so bear with me. But in- they're like, "That's very interesting. You know, what are the other customers that are also utilizing the solution?"

Harj Taggar 18:36 Yeah.

Igor Marinelli 18:37 Uh, and can I talk with them? So it's that whole dance. When you don't have that, generally the, the, the next idea that come in is trial. So since you've lacked a way to provide the social proof, we now need to try the solution. That doesn't happen with me, right? I, I just like, you know, i- within sort of the way that we've designed our technology, uh, I think, you know, going the trial route, imagine like what I've always told my customers, "Hey, you were there for how long? The last 50 years you're producing the same thing, and you're probably gonna produce the same thing for the next 50." Right? So it is not really the trial that also moves the needle, but you have to gather commitment from one side in order to also move resources on the other side. So I, I much rather prefer the term pilot or, you know, paid trial if, if you may, but it has to have commitment from both angles-

Harj Taggar 19:39 Yeah

Igor Marinelli 19:39 ... because it's not a lack of money on their side.

Harj Taggar 19:42 Yeah.

Igor Marinelli 19:42 Right? It is truly just, you know, is this something that, that, uh, i- it is, it, it will work as intended and it is applicable for my situation. So we definitely leverage a lot of those paid pilot programs in the beginning, and that was sort of like gave us room to breathe on, "Look, we're not gonna do anything for free. This is expensive, right? And you also have a huge problem to solve," and that's how we gather that first-

Harj Taggar 20:08 And it clearly-

Igor Marinelli 20:08 ... back and forth

Harj Taggar 20:08 ... went successfully of course, but did, did, how did you define the conditions for success when you're doing those pi- um, paid pilot conversations? Like, how, how specific were you about this is kind of what we can deliver and if we do this, um, this pilot will, you know, will convert or, or renew or, you know,

Harj Taggar 20:24 whatever the-

Igor Marinelli 20:24 Very specific. You have to have success criterias and, and, and sort of like mutual act- action plans that if the customer can even sign off on that, right? It's a shared document between you and them. This is e- absolutely crucial because if you cannot get the answer from them on the questions that you're asking, for example, "Harj, we came in here, we're going to start with 30% of your critical assets. Like, if this works given this such performance criteria and SLA condition, and also the look and feel of the platform, it has to be an amazing experience for you. But given these conditions are met, where are we going to?" Right? "Are we now a long-term partner? Are we now, you know, going from 30% to 60% to 90% deployment?" And customers that are unable to answer that, that has always been a yellow flag for me. And reality is, you know, we've let a lot of those go too, right? And, and, and you can't really have this type of customer when you're just starting your company too-

Harj Taggar 21:31 Yeah

Igor Marinelli 21:31 ... because it can basically kill your company, right? So like it, it, it will kind of hurt, you know, you know, salespeople will say, "Oh my God," but, you know, you have to hug that pipeline and not let them go.

Harj Taggar 21:40 Yeah.

Igor Marinelli 21:41 But you, you truly have to mo-

Harj Taggar 21:43 Yeah

Igor Marinelli 21:43 ... to, to move a, a huge bucket to lost so that you can focus on those that are like giving you commitment, right? Like, so not maybe not as you expected or not ideal, but like, okay, you know, if this works and so, so and so and so and so, I will commit to going from 30% to 50% deployment.

Harj Taggar 22:04 Yeah.

Igor Marinelli 22:04 That's something, right? So otherwise they're just shopping around, they are some sort of sommelier of-

Harj Taggar 22:11 Yeah

Igor Marinelli 22:11 ... you know, good enterprise solutions-

Harj Taggar 22:13 Yeah

Igor Marinelli 22:13 ... and you don't need those.

Harj Taggar 22:14 And a real waste of, yeah, like bad, bad-

Igor Marinelli 22:16 Mm-hmm

Harj Taggar 22:16 ... waste of founder time.

Igor Marinelli 22:17 Mm-hmm.

Harj Taggar 22:18 If we sort of fast-forward, like you actually, this period, you're getting your first customer and you're growing the business, this was early days. This was all pre-AI, and then AI kind of came in, or at least pre-AI in the sense of ChatGPT. Um, AI sort of was like, you know, this massive shock. It seems to have been a huge boost for your business. Maybe just talk us through a little bit about, um, what changed as these models started taking off and how did that sort of affect your business and may- maybe even kind of how did you adapt to it as a, as a founder of a, of a growing company?

Igor Marinelli 22:47 When GPT more specifically came out and we started thinking how do we go from a complete deterministic set of artificial intelligence that we have at that time, it was, it was like neural network technology in the same way that radiologists use, right? So like it was, you know, our, we, we turned our key users from general practitioners to radiologists, right? Like that they were doing this, you know, it was almost like, you know, now we can tell what's, what's wrong with the x- uh, wi- within the x-ray for them.

Harj Taggar 23:22 Yeah.

Igor Marinelli 23:22 What AI allowed was truly the aha moment. It was like always the missing link of, okay, you know, now we are that prescriptive, but how do we intake their experience, their context, their feedback, their knowledge of the machine that we don't have into our solution? So we, we jumped very, very quickly on that. Al- almost like also I guess an idealistic attempt to be the autopilot for our customers-

Harj Taggar 23:53 Mm-hmm

Igor Marinelli 23:54 ... to very quickly shifting to a co-pilot or essentially like, you know, going from the position of we are here for you, and, and, and this is the amazing stuff that you can do with our technology and, and then, you know, our customers went from being a part of it to really owning the technology, I think. So then quickly we started like buying NVIDIA GPUs.

Harj Taggar 24:22 Ah, amazing.

Igor Marinelli 24:23 So yeah, we have a bunch of those. Started- Doing, you know, different tests of, like, pre-training, post-training. Because reality is, like, we have this huge ontology, which in our case is very valuable and it's very domain specific, right? It's, it's 350,000 assets that we're powering, uh-

Harj Taggar 24:41 Which is ingesting data from every-

Igor Marinelli 24:42 Yes

Harj Taggar 24:42 ... single day.

Igor Marinelli 24:43 Yeah, you know, in real time right now, right? And then, you know, the billions of files that we have then from the technical specifications from this ontology that our customers are uploading, right? And, and it's within their, their service orders, is within their, their, um, you know, different data sheet and applications and digital twins of how, how is that real operation and how is it, how is it working. I remember, you know, seeing a lot of Balancer's technology too, and, and the insight came when they also started using Nemotron, which was, uh, basically NVIDIA's, uh, open source model. And, and, you know, they were a little bit mocked in the beginning that, like, it's kind of not possible that you use such a small model with their ontology and you produce better results than Claude Fable 5. But it's truly fascinating when you get that and you start a post-training on your specific ontol- ontology for specific tasks within your domain and what you can do with it. So, and then, you know, reality is after we've done that, it opened different sort of, uh, business models for us as well right now.

Harj Taggar 25:54 Let's talk, actually, so

Harj Taggar 25:55 let's go into that. But yeah, I mean, you're, you're also post-training on top of Nemotron. Um, w- what are the results you see, um, that your customers couldn't get if they were, um, trying to just, like, you know, just calling, or maybe even you couldn't get-

Igor Marinelli 26:08 Yeah

Harj Taggar 26:08 ... but if you were just calling, like, um, the latest, um, Fable or Astra or one of, one of the frontier models?

Igor Marinelli 26:14 Yeah, so the first thing is if you're trying to do the same thing with the latest, you know, state-of-the-art, so get Fable 5 right now. They're, you know, we've run, you know, different benchmark, um, to always try to assess can they do the same thing, and if yes, at what cost?

Harj Taggar 26:34 Yeah.

Igor Marinelli 26:35 So, so there's two dimensions to this, right? And reality is Fable 5 is the closest, right? So, you know, basically can execute 85% of what our models can, reach to the same conclusion. But then the token consumption, it is basically 20 to 100x still to get to that same conclusion, right?

Harj Taggar 27:00 Which is huge for you guys, 'cause the whole point of sort of like, um, you need to know the, like, the second before a machine has any issues, um, which means you need to be sort of consuming data, like, 24/7, like, every minute of the day, right?

Igor Marinelli 27:12 Exactly.

Harj Taggar 27:12 And so, like, cost must be a huge, huge factor.

Igor Marinelli 27:14 It-- And, and then it comes the second portion too, which is, you know, after training, post-training with, with the open source models, the hardness is very important as well.

Harj Taggar 27:25 Mm-hmm.

Igor Marinelli 27:25 Because there are different, you know, technologies that you can do instead of, like, just reasoning and reasoning and reasoning with tokens. You can use this array of tools which are still deterministic, right? Because you are looking at something that is within your domain and, and, and you don't need a lot of guessing game just to get to 50% of it. But what, what we do then, we also then test with, like, humans, right? Which I, I think it's important. So we always run the test with, like, get, you know, the best PhD in mechanical engineering, and they're looking at this asset data, they're looking at this ultra- ultrasonic thermography picture and, and, like, everything that's available that's also available for our model. They're looking at everything and trying to assess what's wrong with the asset, right? So, so we do this reverse test as well. And, you know, and it's sort of like this, this step of, of where, where, like, a PhD in mechanical engi- engineering, it's generally 40, 45% right with all the time in the world, no time expiration there. Then a Claude Fable 5 can get to 85% with, like, 100 times the cost of what we can get to that, you know, 100% essentially. So in the moment that they launch an- an, uh, you know, a next model, we're also retraining on a new model of the open source model and so on and so forth. So it is compounding and, and I think that that's, that's the one piece. The second piece is we also have a lot of customers that are on premise.

Harj Taggar 28:59 Hmm.

Igor Marinelli 29:00 So-

Harj Taggar 29:00 Oh, yeah, they don't want the data leaving their-

Igor Marinelli 29:02 Right

Harj Taggar 29:02 ... facilities or their cloud or-

Igor Marinelli 29:04 So, so it, it, again, it is kind of like, you know, uh, not popular to, to say that, you know, we deploy on premise, but on premise is one of the most kind of like beautiful things that, that, that, that, you know, it is still available for everybody to... And it is modern still. Um, because even if it's in, even if it's, like, in the cloud, so it could be in my cloud, it could be in the customer's cloud, or it could be on premise, on premise, like bare metal on premise. The best design for this sort of enterprise application systems, you know, the really, the traditional SaaS, which is like, you know, your data's here and it's shared with everybody and so on, it goes just up to a certain point. Enterprise buyers, they want to say, "Hey, give me the latest model that you have-"

Harj Taggar 29:50 Yeah

Igor Marinelli 29:51 ... "that you've already trained with everybody's data," right?

Harj Taggar 29:54 Yeah.

Igor Marinelli 29:54 "Which is anonymized, et cetera, et cetera. But from here on out, I want this to only to compound with my own data."

Harj Taggar 30:01 Yeah.

Igor Marinelli 30:01 And they are huge-

Harj Taggar 30:03 Yeah

Igor Marinelli 30:03 ... huge applications and so on.

Harj Taggar 30:03 Owning, companies want to own their own intelligence. That's like a, a huge issue.

Igor Marinelli 30:07 And I-- Wait, and by the way, the, you know, we're telling, you know, this, you know, this very cool story right now, but I remember one of the decisions very clearly with our CTO at that time was just basically like, this ontology that we're compounding, why are we sending to Anthropic? Why are we sending to OpenAI? You know? And I remember even going to their privacy page and like, you know, can I really switch off? And the reality is like there isn't a way that you can be sure that like that PDF that you're sending is not gonna be used for their own training and so on and so forth. So you also want, you know, that domain to stay there.

Harj Taggar 30:46 Since you have this really unique position of, um, AI deployed like in the real world, in real sort of facilities, um, uh, and you're at the forefront of, of that, sort of where do you still see kind of gaps maybe in AI, and where are you sort of excited for, um, uh, for it to improve or for you guys to improve it so you can deliver even more to your customers?

Igor Marinelli 31:09 It is changing the way people work, and if even our employees, now they h- they have this army of agents-

Harj Taggar 31:19 Mm-hmm

Igor Marinelli 31:19 ... doing things for them, right? Bookkeeping, accounting and so on. Even, even like, you know, tech workers truly, right, like that were als- sort of like born and raised and, y- you know, already with, with internet in their hands, are sort of like, "Okay, I have all this power in the world, and how do I use it?" And you want to a- actually measure, have I truly become more and more and more and more productive with that? So I think it's like the, the f- the format that this is still embedded, uh, I think there's a lot of improvements that are still going to happen there, right? What is the best way to interact with it? Is it a bot with-

Harj Taggar 32:03 Yeah

Igor Marinelli 32:03 ... a, a, a, a, a, a face of a real person?

Harj Taggar 32:07 Yeah. Yeah.

Igor Marinelli 32:07 Is it embedded in the systems? And, and so I think once we actually solve the mediums which we're delivering this information-

Harj Taggar 32:16 Mm-hmm

Igor Marinelli 32:16 ... things are going to get much more clear. In the case of our customers, what's in- what's interesting about this is like there is labor shortage in our market. In manufacturing in general, it is already h- like extremely hard to hire.

Harj Taggar 32:30 Mm-hmm.

Igor Marinelli 32:31 It's not exactly hard to hire for startups the way it is hard to hire for manufacturing.

Harj Taggar 32:36 Yeah.

Igor Marinelli 32:36 Right? So that thing is coming from them, and it's coming from them like with, you know, so much ease because what this is allowing them to do, it's they could already not hire. They're also not firing because those employees are extremely crucial. So but now one employee, you know, what, what I was talking with one of our customers, can this one employee that was just like regional to this one facility also control and operate that other facility-

Harj Taggar 33:05 Yeah

Igor Marinelli 33:05 ... that he's not, you know, locally there. And then that is like the next leap that this technology is allowing them to do. For our own company, what's interesting is it's, you know, the, the balance between sales and engineering, right? So like I, I didn't think I would see, you know, engineers face the, um, uh, the result of their work coming at them more aggressively-

Harj Taggar 33:39 [laughs] Yeah

Igor Marinelli 33:39 ... than salespeople, right?

Harj Taggar 33:40 Yeah.

Igor Marinelli 33:40 'Cause everybody's thought about like, look, you know, we're now going to buy from these AI bots.

Harj Taggar 33:45 Yeah.

Igor Marinelli 33:46 And it's just not happening, right? Like, engineers, on the scale of productivity, are getting 100 to 1,000 more productive.

Harj Taggar 33:54 Mm-hmm.

Igor Marinelli 33:55 I don't recall the last time we hired an engineer because we keep reviewing that, the headcount plan and keep pushing it forward. But sales, we're still hiring them. Maybe they got 30% more productive, but customers want to buy from people.

Harj Taggar 34:08 Yeah. That makes sense.

Igor Marinelli 34:09 And, and, and, and that transaction is also-- the nature of that transaction is also not changing, but your back office is changing.

Harj Taggar 34:16 Mm-hmm. Um, I mean, something else I'm thinking about now as you're, as you're talking is you, um, you've just been through this big boost from this transformation, um, GBT transform models, um, uh, coming out. At YC it seems to us like there's, um, a lot of excitement around robots. Um, uh, and so, you know, like AI in the real world definitely becomes more exciting. I'm curious, like in your world working with these sort of like asset-heavy, um, industries and customers, as, as the sort of more, um, yeah, robots enter the space and there's more automation, there's more like maybe smart machinery is a way of thinking about it, like how, how do you see that sort of impacting your business and, um, uh, and sort of like the opportunities that opens up for you?

Igor Marinelli 35:04 We're very excited with, with robotics, right? I think what companies are solving in the robotics space, uh, in, you, you know, in our AI labs, we have the Boston Dynamics-

Harj Taggar 35:17 Uh-huh

Igor Marinelli 35:17 ... quadrupeds and-

Harj Taggar 35:19 Uh, in your own, like the AI lab at Tractian-

Igor Marinelli 35:21 Yes

Harj Taggar 35:21 ... you're working on this kind of stuff?

Igor Marinelli 35:22 Yes.

Harj Taggar 35:22 That's cool.

Igor Marinelli 35:23 We're working in dispatching then automatically in our own model, controlling, you know, their own h- hardware essentially and what they're going to look at. Again, we have the ontology, so we know exactly what they need to scan in that specific asset.

Harj Taggar 35:36 That's very cool.

Igor Marinelli 35:36 So this future is happening at a lab, you know, scenario already. The only reason it's not happening at the customer, they're just not buying it.

Harj Taggar 35:45 Yeah.

Igor Marinelli 35:46 It's the price tag still, right? Like, you know, we bought, you know, the, the quadrupeds, like the Spots, at 100K every, every, um, every dog, and then we put this bunch of instruments on top, uh, which are essentially, you know, the traction gear on top. And then by the time it's a final product, y- you get to a range of like 200 to 300K.

Harj Taggar 36:07 Mm-hmm.

Igor Marinelli 36:08 So either there's going to be, you know, a, a more exponential, you know, price, uh, decrease in, in, in, in what we're seeing in the quadruped w- world, or which we're also experimenting, drone inspection, which is- You know, in its nature it's cheaper. It can carry just one gear, but at the same time, like, it, it's nimble, it's fast. Um, and, and all of a sudden all of those barriers are also, like, of just transporting the thing are also out, right? So I think we're very, very bullish on that. And the thing about, the thing about operations in general, right? So that, this is not new. This has always been happening. So you get into any, [lips smack] um, um, physical, uh, plant right now, what you're going to see is an array of different, um, equipment providers.

Harj Taggar 37:08 Mm-hmm.

Igor Marinelli 37:08 So you're going to have Siemens, Bosch, Schneider, which are not very good at talking to each other, right, by design. And then you're going to see also different robotic arms brands as well. And you're going to see different ages as well, right? So you can have an HVAC from 24 years ago, 15 years ago, and five years ago that was manufactured, and all of them, they have this, you know, they have different protocols and, and different things that they carry, right? So, like, we've got specialized at homogenizing all of that into one single platform, right? And hardware is the answer as well for that. But essentially what we do, no matter if you have shadow spots that are not covered, no matter if you have a machine that's partially covered or a machine that's like, "Oh, this is like the Ferrari of, of robots," right? But at the same time, it needs to be in that same place that you have your other equipment. So that is where we believe that we play like a very, very sort of, you know, uh, wild card mentality into the space.

Harj Taggar 38:18 When you think about sort of your moats or your advantages in, in this world, is it, um, is this with the fact that you've already got these relationships and your hardware's installed, um, at all of these customers and these facilities, or is it the, the data you've collected from that? Or m- maybe it, it's probably, obviously it's, it's sort of both of these things, but I'm, I'm just curious how you think about that. 'Cause, 'cause you're in such a unique position, like physical AI is really just about sort of like starting to have its moment and, and take off, and so I'm just curious how you think about that sort of moat for, for where you are.

Igor Marinelli 38:49 Is it that we know the best on how to operate and balance assets in all the genre that's there and, and, and, and available in, in, um, you know, plants across the world, right? So our models are that close from operating the asset, right? 'Cause if you think about, if you think about what we do-

Harj Taggar 39:20 Yeah, interesting. Yeah

Igor Marinelli 39:20 ... essentially in its nature, it's ins- ensuring capacity in the domain of, like, if it's a transformer, we're ensuring capacity of electricity. If it's a chiller, we're ensuring capacity of chilled water. If it's, you know, c- uh, an air compressor of compressed air, if it's a gas compressor. Uh, so, so, like, w- what we are ensuring essentially is the consumables that are, that are powering something. Those consume- consumables are then going to cool a GPU, and that is out of our domain. But when they're back, this is within our domain.

Harj Taggar 40:00 Hmm.

Igor Marinelli 40:00 Right? So we're not messing with the GPUs of a data center. That is proprietary technology of someone else. But they need someone for that infrastructure of cooling to go back and forth, right? So for as long as we know what is the perfect operating condition of that-

Harj Taggar 40:19 Hmm

Igor Marinelli 40:19 ... then we can put that on autopilot. And, and then we can do that in, at a s- at scale as well. So, and fut- futurally speaking as well, then dispatch a human, a robot, to go there and then correct what's wrong ahead of time so that there is no unplanned downtime in the foreseeable future, essentially.

Harj Taggar 40:42 Yeah, that's interesting. I hadn't thought of it that way, actually. So yeah, you basically, you start by sort of just like logging and monitoring and knowing everything that's going on, but then as the technology improves, it does more than just sort of alert you when something's wrong. It knows what to do, and then, um, yeah, in the future it can actually just like go in and fix the thing and actually operate the, the entire facility, and that's kind of exciting.

Igor Marinelli 41:03 We know to the level of, you know, in specific air compressors, if you swap a bearing from manufacturer A to manufacturer B, you're going to get more compressed air-

Harj Taggar 41:17 Right

Igor Marinelli 41:17 ... with the same electricity, right? So, so we know, you know, to the level of the mechanical design of the asset, to the electrical design and the operational design.

Harj Taggar 41:26 Uh, Igor, congratulations on closing the Series D. That's a, that's a huge milestone. Um-

Igor Marinelli 41:30 LFG.

Harj Taggar 41:31 [laughs]

Harj Taggar 41:32 Yeah. Let's do it. Um, um, uh, something I'm just curious about, um, so much of the audience here are aspiring founders. How does fundraising change as you go from, you know, you know, getting into YC, raising a seed round, a Series A?

Igor Marinelli 41:45 It gets harder.

Harj Taggar 41:46 Yeah. [laughs] Yeah, really? Interesting. Okay. Tell, tell us more. Like what, what [laughs] What's, how does it feel different?

Igor Marinelli 41:52 There's so many different angles that you can look at a company, right? Like, you know, we crossed $100 million of ARR.

Harj Taggar 41:58 Congratulations.

Igor Marinelli 42:00 It's, it's... Well, it feels kind of a big number, right? Like-

Harj Taggar 42:02 [laughs] It's really

Igor Marinelli 42:02 ... we've done it in seven years.

Harj Taggar 42:04 Yeah.

Igor Marinelli 42:04 But you know, your new companies now are doing it like, you know, doing it in weeks.

Harj Taggar 42:08 Annualize. Annualize.

Igor Marinelli 42:09 They're doing it in weeks.

Harj Taggar 42:10 [laughs] Definitely.

Igor Marinelli 42:12 What am I doing wrong, you know?

Harj Taggar 42:13 [laughs]

Igor Marinelli 42:14 So, you know, but, you know, jokes apart, like we're extremely happy with that, just, just with the nature of the way we distribute products, the way that we get to customers. You know, we have this, this field sales combined with like a way that a tech company operates, which makes Tractian very unique in that sense. But, you know, fundraising- It goes through a, a process and a journey of, well, in the beginning it's about, about the founders, it's about the market.

Harj Taggar 42:44 Yeah.

Igor Marinelli 42:44 Right? And execution, and you tell me the, the third thing, right? So this is your [laughs] specialty essentially. But so, you know, definitely, like, I look at my first angel investors, and I asked them, "What was wrong with you guys?"

Harj Taggar 43:01 [laughs] Yeah.

Igor Marinelli 43:02 Like, when you look at first versions of Pitch Deck, I would have never invested.

Harj Taggar 43:07 [laughs]

Igor Marinelli 43:07 Like, you know, you guys are completely out of your mind. But you were feeling something special about the founder. You, you were, um, in, in, in s- in, in certain ways as well truly, like, also fact-checking, can this person walk the talk? Can they hit their milestones? And, and, and sort of like, sort of like a very intuitive journey-

Harj Taggar 43:27 Mm

Igor Marinelli 43:27 ... on an individual level.

Harj Taggar 43:29 Right.

Igor Marinelli 43:30 And then as the company grow, you know, we have to keep reminding investors, "Hey, it is still about the people that's here," right? And that's, that's how we've been getting there, and that's how we will continue to get there. So otherwise they will definitely, definitely be like, you know, spreadsheet investors, like it is characteristic of late-stage investors, right? Like, you know, spreadsheets after spreadsheets af- after financial calls. And, and yes, great numbers tell a story, but, you know, we're still building this, and we still, we don't plan to go anywhere, and we plan to build this for the next, you know, 15 years plus. Um, which is my mentality, right? It's, it has always been long term. What's characteristic for me, I think, is just I've been more of a, a founder of quality over quantity. Definitely has, always want to make sure that, like, quality is in there, but when it's in there, then we double down, uh, on that. So it, it's changing. It's different. It's different every time. I think, you know, it, it doesn't compare so with the, with the feeling of receiving that 125K.

Harj Taggar 44:42 [laughs] Nice, okay.

Igor Marinelli 44:42 That was just like-

Harj Taggar 44:44 [laughs]

Igor Marinelli 44:45 It's just like, you know, nobody can beat that, right? But, uh, it is also special in its own ways.

Harj Taggar 44:51 Cool. And so now, you know, the company is worth, uh, $1.4 billion. Um, what kind of has to happen or what has to be true over the next few years for it to, for it to 10X, for it to be worth $14 billion or, or more?

Igor Marinelli 45:03 What is happening right now that we are already seeing when you get to that certain size, there is an interesting power law that starts happening within, you know, the funds that you're bringing and the customers and the company again, and, and you have like a flywheel effect essentially. That's what we're feeling right now, right? Like we have, you know, funds that invest in Tractian that they also own and operate data centers-

Harj Taggar 45:30 Oh, interesting

Igor Marinelli 45:30 ... that we are in, and then it comes back here. So it's just like, so it's that serendipity that ends up happening. In the beginning it felt like an overkill, right? And customers would, would also say that in different ways. Like they would quickly recognize, it didn't take long for them to say, "Traction is so, so, so far ahead from the nearest competitor." Uh, which in underserved markets tend to be the case, right? When we have 100 patents, and we keep publishing patents, and it felt like, is it an overkill? Because we were essentially talking with customers and then building products that we ourselves, if we were there, we would like to use, right? Always with that originality of like, we can solve problems one way, which is solving it, but then the other way, which is like solving it for us if we were there. So we're always doing that B, and which kind of felt like, like an overkill. But now as we've also, we're able to establish, you know, multimillion-dollar contracts and, and talking with, with, with, with customers that have also that high tech mentality. Now we're like seeing that, oh, that's very interesting because of the way that they choose their next technology as well. It's not only based on what's available but like based on where they're going and if you can meet their velocity too.

Harj Taggar 46:58 Cool. And then maybe sort of just to close out and bring this full circle, you know, this journey, a lot of it started with, um, your personal story, your father, his experience as a, as a plant manager. When you sort of look forward, if Tractian sort of, as it fulfills this vision and it, and it works and it scales, how do the lives and the jobs of your, of your customers change over, over the future? Like what, what becomes new for them?

Igor Marinelli 47:23 I think what we're doing, we're compressing time in such a way that... Think about this. For majority of our customers, the first time ever that they're using AI is when they're using our solutions.

Harj Taggar 47:40 Mm.

Igor Marinelli 47:40 So just, you think about the magical moment of a programmer looking at cursor and just like, "Wow," you know, "This thing was... What is this?" Right? And so we get to almost double stack that magic and, and compress time in a way of like, "I had no idea this ex- this even existed, and, and this is what I am capable of doing as an individual." So there is like the individual sense of they get to come home at 5:00 PM, which for me is key for my customers, right? Like they have that predictability, which my father didn't have, like weekends at work, you know, burning the midnight oil, all of that, uh, which is always related to a piece of something that's failing that they didn't have any knowledge of. And as a bonus as well, we get to save billions of dollars, um, in meas- measurable economic impact for our customers. Like I have in my email signature a, you know, savings counter of Tractian to customers-

Harj Taggar 48:52 Oh, wow

Igor Marinelli 48:52 ... which is right now at $1.2 billion.

Harj Taggar 48:54 Huh, wow.

Igor Marinelli 48:55 Like our dream is to get to $1.4 trillion.

Harj Taggar 48:57 [laughs] Nice.

Igor Marinelli 48:58 It's just 1000X.

Harj Taggar 48:59 [laughs]

Igor Marinelli 49:00 Right? So I love the fact that like, that we can measure it so simple, and so you have like all this like incentives aligning from the personal motivation of getting better conditions for frontline workers to also the economical motivations, which is this has to produce, and it is producing. It is producing in like some customers reporting, for example, you know, one gigawatt of savings in energy, millions of, of, of gallons of waters that were saved. So like it is measurable at the same time at an individual too. It's very satisfying.

Harj Taggar 49:36 Cool. Okay. Well, that's a, um, empowering vision for the future. I think that's all we have time for today, Igor. Thanks so much for, uh, being here with us, sharing the story, and congratulations on the round, and we're excited to see Tractian go from strength to strength.

Igor Marinelli 49:49 Thanks for inviting me. [outro music]