The Art of Chunking with Dr. Alexander Kihm | Ep. 59

Speaker A: Okay, let's build a directory of lawyers online. And I said, there's already three. I don't build copycats, stupid. The first nine all got rich, but I was arrogant by design, so, uh, nah, it's already there. I don't build it.

Speaker B: Hey friends, this is podcast ruined by Software Engineer on M. This episode I get to talk to Dr. Alexander Keen, founder and CEO of Puma AI. In the world of large language models, also known as LLMs and Retrieval Augmented Generation, also known as RAG, PUMA is, uh, a chunking technology that consumes large libraries of data, JSON PDFs, images, chunks it with its proprietary solution and outputs a context rich data that can be consumed with higher fidelity and accuracy in any applications. Simply put, Puma is a new type of engine that consumes less gas and outputs cleaner, more structured power that could be used in whatever vehicle you could think of. If you enjoy podcast written by a software engineer, don't forget to hit the follow button on your podcast app so you don't miss out on the new episodes and you can pick up podcast merch@perry2.com shop. Enjoy the episode. Welcome back to another episode of podcast written by Software Engineer. I'm your host Perry, and today with me is Dr. Alexander Keem, founder and CEO of Puma AI. Alex, how are you today?

Speaker A: Hi, happy to be here today. Thank you.

Speaker B: I'm going to preface this. I'm going to be so done this episode. We're talking about chunking, we're talking about AI, we're talking about rags, we're talking about LLMs, M. These are all the great keywords, but you've built upon this today. But your journey to getting into that space is obviously so much more interesting as well. So we're going to be diving into all those kind of bits. But before I dive too much into it, who is Dr. Alexander Keem and what is Puma AI to begin with?

Speaker A: I'm Alex. Uh, I'm an engineer. Like very classic. Starting with the LEGO bricks at three, then first computer at five, disassemble it at six, um, try to tinker with it at seven. At eight. I had to reassemble this because I think too much like, yeah, ah, engineer, you get the gist. Um, I'm also an economist, which is a bit strange because after an engineering degree, why would you do your PhD in economics? But it's basically big data econometrics, so the common denominator is code. I'm actually not good with my hands. I have two left hand and, uh, like The DIY stuff at home is my wife. Um, so, yeah, I'm the coder. I used to make this fun when we meet people in a bar and then they ask, what do you do for a living? I'm like, yeah, I'm a text butcher. Um, I'm a text butcher. We actually butcher texts so your AI can consume them. Because AI doesn't eat the whole animal at once. So we make the steaks, the sausages, and everything else.

Speaker B: Because, I mean, like, there's so much work behind preparing what we could present to end users at the end of the day. Right. So that's really what the magic sauce really is behind Puma AI, which we can't wait to discover. But sometimes what I do, like, for a lot of people to, like, kind of just like, situate ourselves and even be more relatable is that we all grew up somewhere. We all have fond memories of the influences around us, the people around us. So basically, where did you grow up? What area of. Which part of the world did you grow up in? What do you remember from that?

Speaker A: Yeah, uh, I grew up in Dusseldorf. It's, uh, well, it's actually, it's. It's not an unimportant city. It's the capital of North Rhine Westphalia, which is the most populous state in Germany. Um, it's in the northwest. It's close to the Netherlands. So back then when all the. All the rage in my youth was like, people driving by bike to the Netherlands, getting some. Some weed and get back, I mean, back then it wasn't legal. I. It's not my thing. But it was always fun, like with the. Okay, where's this guy? Ah, he just drove to Fendo again. Um, so. So it was really like, part of the youth also. That's one of the reasons why, uh, the state was generally pretty liberal and also pretty like one big city. I mean, it's 25 million people basically living in city to city. So it could be like one city. So very used to, like, urban, Urban sprawl. Um, I'm. I'm a city kid. Um, and Dusseldorf is, uh, quite a fancy city. I didn't realize it as a kid, only later when I, When I, you know, saw the rest of the world. So I've said I had this quite protected, uh, upbringing. Like, oh, everything is nice. Everyone has two cars and whatnot and super easy. But you had to just cross the bridge and you were in the middle of the action. So, um, that was kind of cool. It's also known for the longest, um, the longest bar in the world, which is basically, there's like, yeah, there's like in the middle of the city, there's like these 10 or 20 bars. And they kind of put their, you know, put their counters so as to. It's one continuous one and it's very well known. And it's like, you know, there's like always stack parties and whatnot. But it's, it's kind of, you know, I knew bars from an age where, you know, many people in the world wouldn't start going out in a bar, let's say. But then Germany is also very little on beer drinking, so that was kind of cool. That's my, that was my background. Um, and my parents are doctors actually. So, uh, I had this like knowledge culture and you know, I wouldn't say academic upbringing because they're also pretty straightforward and honest and they're also very supportive. I'm from a sports family actually. Almost everyone around me is high ranking in tennis. My sister is even still coaching, my brother's still coaching. My parents, Both in their 70s, they're still playing in high leagues. Um, but yeah, I'm the blind guy. So I'm the guy literally with you Remember these kids that had like one of their eyes, you know, plastered over, so the other eye gets and so on and so on. So when I was 4, they cut part of my eye muscles. So I'm not cross sighted anymore. I'm very thankful for that. Good job, doctors. But that obviously makes for a little bit less, uh, coordination when you only see two or three dimensions. So I'm always used to, I'm the small one, I'm the blind one. But give me a computer and I can help you. Um, and uh, so yeah, that was a bit like my role there. I'm the eldest son, fair enough. And have two siblings. Um, so there was also early on a bit of this. Yeah, Learn to manage. Learn to manage a subunit of an organization. Like I remember when my parents came, like always. Yeah, but it's such a mess. The kids have to clean up and so on. And at one point I said, guys, instead of abstractly rambling to the room, what the kids should do. Are you actually talking to me? I think yes, because I'm the only one already speaking here. If that's the case, can I please. I will take responsibility, but I will also manage. And you don't, you know, break. Uh, you don't breach hierarchy, you don't sanction, you don't reward. I'm doing it. And in the End you say if you're happy or not literally. Because it pissed me off at one point. And they say, okay, deal. So I manage my siblings. I got coffee at 14. There was the deal. And beer at 16. That was the other deal. Um, and. But I was managing the family and I was making breakfast every morning and I was, you know, keeping my parents, uh, like, you know, random management away from my, from my siblings, but also my siblings mess away from my parents. And later I thought like, oh God, was it overbearing? And I had like a lot of, you know, reflections about it. And in the end all, all people involved told me, no, no man, that was brilliant. And uh, uh, we all love you for that. No, that was, that really meant something to me. I thought like, oh no, did I, did I go like for the, you know, the shit rolls downhill kind of bad manager type. But in the end that, that made like maybe that was, that was an important part in me not just you know, being the nerd with the computer, but I actually like to lead and uh, and I actually like to take the responsibility and, and say, okay, there's a problem. You can do something, you can do something. There's like, you want this there or that? Uh, let me manage so we can all get along and get this thing running. That's. I, I only found this out later when they told me, but apparently there was a big part.

Speaker B: This show really just gives me the opportunity to see other people's upbringing, other people's environment, if anything. And also, I mean we've barely been on this show for like a couple minutes and you've already taught me like six different things. If anything short is that you're not short of like interesting facts like the fact that you could actually correct cross eyedness like that. I actually didn't know about that. When you're seeing in terms of like the environment, being able to like literally cross country, not the running cross country, but being able to like be exposed to other countries culture. The interesting part is that like for Americans it's like you drive seven hours and you still within the same state. The culture is very much the same and everything but in Europe is like you drive three hours and next thing you know you have probably having paella on a beach instead of like wine like in France and things. So that exposure, it's not just about like the culture, but I guess a way of living. And you're seeing, maybe I'm being dumb here, but like at that time, if you were crossing between um, the Netherlands and, and Germany. Do they still require you to like wave your passport or like how does that work?

Speaker A: It's actually funny. You know, I myself had to let there recently because of like all the you know, mounting, um, the, the border back up, um, discussions all over the world. I had to look up like when was Schengen? And it was I think 91 or something. So no it was uh, it wasn't like a border fence or something but it was like the semi open border. So whenever you could, you could just cross. But then every ex or, or every X car or every thing a bit suspicious. This is why in the end my friends went with I think racing bicycles and like full, full dress and so on. So they were like the super sportiest and in their back they had like all the weed. It was super funny. It was always the joke like someone came like oh wow, you're from the Olympics. No, I just came from Holland transporting weed.

Speaker B: Isn't it the same thing?

Speaker A: Yeah, exactly, exactly, exactly, exactly. That was like the running gag there um, now. But this whole um, now Europe was already pretty open. I mean you know, maybe like they sometimes wave when you had like a lot in your trunk. But uh, generally I grew up for example like over my childhood I realized I spent a whole year in Italy because we went there so often for always like long stretches of vacation, three weeks per year, always same, always same area, always same hotel. And I liked it because my parents, they liked like you know, their, the rituals and their, their you know, some, some things never changed and I kind of, I now kind of like that. Um, and but in the end that made me like uh, being fluent in gastronomic Italian and like all day Italian at the age of I don't know, eight or something. And I, I was the bingo translator in a beach club. It was so funny. You know the, the guy telling the number in, in Italian and then me like uh, with my little Italian like translating it to the German audience. That was super funny like stuff like that. So the early exposure to, to other cultures is always a thing like I loved. And that's why also you know I'm, I mean now Berlin is like the total opposite, uh geographically but also culturally from, from uh, you know, the, the west or especially the south where I did my studies. But then also you know I love. I did the spell in Reunion, like the island between Madagascar and Mauritius, uh, where I learned French and Creole. And then later I went to Chile. So I'm, I'm really, you know, I wouldn't consider myself like a Global citizen or cosmopolitan or whatever you want to talk it. But I'm very open and I love seeing new things. Like, I went to Japan two years ago for the first time and I, I soaked it in and I always, you know, I don't, I don't consume travel anymore. I don't book like some, you know, tour to acts. It's more like I have a friend in Tokyo. I wait until I have four weeks and then I visit this friend and I, you know, really get to know stuff. Get to know the country or, you know, this is why, for example, I love doing business in other countries. Not to do business, but to, you know, I go somewhere on a mission and I don't go there to consume the place. It's like, uh, you know, give me, give me, give me something to do. Could be visiting a friend, could be something. But not like, hi, I'm here because, um, I'm here, you know, like, why? And I, I want to be part of the product and not like the consumer. Like, that's the idea.

Speaker B: Similar. Like you. When I go to a town, I get obsessed with like, how local lives, they, they all have to deal with the dmv, their version of a dmv. DMV is basically Department of Motor Vehicle. Like every city, every country has something similar. And sometimes I'm so dumb, I'd be like in the middle of like on a Bali. I'm like, how do people get their license? Yeah, like, how does their like administration? Like, do they line up? Do they like send an email? Did they like send a form? Like, this is me in the middle

Speaker A: of like, I totally get it. I have the same with, with, with number plates. I'm obsessed with number plates. Like, okay, how does it work? Like, what do you put on number plate? Is that your name or is that the registration age? Or is like, uh, how. How do vanity plates work? And I don't know, exactly the same. My wife realized also when, when we go somewhere new that I'm always stuck at the, you know, the realtors, like the, the real estate agencies. Not because I, you know, I mean, imagine I'm. I'm not exactly. Well, first, I'm not rich. But second, I'm not the type, like, oh, cool, let's buy something. But I always want to understand the market. Like, the first thing I do, like, I enter a new city and if there's a real estate agency, I'm just like, huh. And then the next one as well. Starting to triangulate how it works. Like, okay, what are they looking at, like, you know, in some countries, energy certificate is half the. Half the. Half the ad. In other countries, it's like, like, measures you cannot understand. Like, qualifies for xyz. And then it's like five numbers. So how much does it cost? Okay, I don't, I don't get it. And then it's like in zone A and Zone B, this is a very good XYZ opportunity. And I don't know, like, I mean, housing is the first thing in, in life. You know, it's, it's, it's. It's like, so interesting how this works in other cities. Um, so I, I always try to get a grip on it quite early. It's like, part of people's life. I like this.

Speaker B: I absolutely love it. Like a lot of people. Okay, this is the last comment before we move on to the actual tech part. Because, like, there's a text show, but, like, when I was saying that, like, you know, people love their museums, people love their landmarks and a new town, like, we're just geek out on, like, random societal, like, topics that everybody deals with in every city, and we're just like, how does that function? Which, which is why. Back to the show of podcasts ruined by a software engineer. Podcast ruined by multiple software engineers today, if anything, you're more. You're a software engineer. As much of a generic engineer as far as I can tell, because we'll get into a bit of your background in terms of how you learned your craft to become where you are today. But one thing I did want to take the opportunity to talk about is that this kind of looking back in the childhood thing, thing is that I do this a lot. And at some point I discovered the technologies that I've used growing up. Like, for people who don't know what's a Tamagotchi, it's like, it's a little, like, pet. It's black and white. It's like, it seems so simple. But looking back at it, was there like a, uh, chip inside? Was it more or less powerful than a Raspberry PI? Like, I'd never actually, like, gone down the question of looking at what technology. So maybe as a more generic question to you, though, we've kind of had a little glimpse of your upbringing in terms of, like, surrounded by siblings, surrounded by a very, uh, interesting neighborhood. If anything, where there's multiple cultures and you're saying it's very athletic, but at the same time, you're like, maybe I'll stick to, like, building stuff. There's a lot of these influences. But what do you remember most in terms of specific toys or technologies or anything like that that you got to play around with? Push buttons, anything like that?

Speaker A: Yeah, I can. Well this, this, it really boils down to two things. And that's LEGO and the PC. Um, and, and Lego, I mean for real, like serious. Uh, you know, I started with uh, with Duplo like when I was very young. I don't know if you remember. Like this is like the um, twice the size LEGO kind of thing in, in Germany it's called Duplo. Um, so basically double. Um, and the thing is I realized, okay, you know, okay, it's obviously that's, that's, that's too rough for me and that you can only like put bricks on top of each other, but they're compatible. So I realized, okay, let's build like the foundation and the rough outlines of the big castle in Duplo. And so we have like already quite some massive thing and then we build on top of that. So you could, you could build something like that looks, it looks a lot more impressive because you had a foundation of something simple. And then you know, you could fine grain it because they were, they were interconnectable. Um, there was like few people actually know that they are compatible. Super funny. Maybe in the US it's easier when they call it mega blocks and then you get it like it's bit a little LEGO mega block. But in Germany and they're marketed as two different products. So maybe that maybe so their parents buy another stuff again. But um, so yeah, Lego it was and also LEGO Technic. But uh, but like the serious stuff like with engines, with even a gearbox and, and whatnot. And I remember like my parents when they said okay, okay, there's snow in front of the door. I mean which is. It was, wasn't rare rare but you know, it wasn't exactly like uh, every year, uh, occurrence they said hey, um, Alex, so you're in charge of uh, getting rid of the snow in front of the door. And I said m. Okay. So I built a truck with a battery and with a grinder for salt and everything. Like okay, I have to automate this thing. I mean it was like 2 meters front door and not, not much to do. But I needed this thing to work automatically. So I built this whole. The Lego. I mean LEGO is not exactly made for outdoor salting use, but who cares? Um, so yeah, I made this truck and then it grounded the salt and it drove in front of the door back and forth a bit like today's room bus. Um, okay, well it rammed a lot more stuff in its way. But you know, since it was only Lego, there was no damage. And I was so proud. Like look, look, it did the job and soldered the door and everything was cool. That was like exactly my stuff. And the other was uh, given that my father, given that they, they rented doctor's cabinet in the, in the home basically, well, downstairs in the house. And that meant um, you know, in terms of uh, the whole depreciation and whatnot. Tax rules. I always say if you want to understand society, look at their tax code. And the tax code said that every X years. My father's computer, he was very innovative and always had like a computer in the, in the, in the doctor's cabinet. It would depreciate. So he said, you know, okay, uh, there wasn't really a used market and he was very early. So I got my first computer so early because it was a depreciated hand me down asset from the doctor's cabinet. So I had like, I think it was a 386, funny enough, my second computer was an 8088 like the. So before 286. And I had them both in my room because I just, you know, I just bring it on, bring it on. Give me the hand, hand me downs until my brother also wants one. So I didn't need a heating anymore. Yeah, we could put the heating out. And I heated the room with my two tower computers. They were as tall as, I mean back then they were taller than me. Now it's like half my size. But that was like, that was my kingdom. Loved it.

Speaker B: Oh, uh, you've answered the question for me. On this show. I always ask what is your first computer? And you've absolutely just blew that out of the park. Windows household, is that correct?

Speaker A: Yes, yes, it's um, Yeah, I would even say DOS household, uh, which is an important differentiation because my father, he resisted uh, the move to Windows, at least for the, for the, for the doctor's business for so long that even, even at the end he had like a full DOS emulation inside, you know, the Windows ecosystem. So yeah, I was really like Terminal Kit, um, so, so DOS for a long time then correct, Windows. And uh, I myself transitioned uh, to Mac, uh, 2010. And ever since I'm hardcore Apple. Like we recently, recently found out with my wife, like here in our household is there's no more CD drive, which was interesting. Just like we found out there's no CD and Um, no smart, um, no smart appliance that is not from Apple. Like laptops, like tv, whatever, whatever, whatever. We, we appled it all. Um, not blind fanboys, I would say. But I always like to say, uh, you have to decide which giant to trust. It must be one, otherwise you're in the DIY hell and we don't have time for that. And the only one I could trust a bit was, would be Apple. So that's like, okay, sell your soul to one guy and that's the guy that sells your hardware and not data. So yes, that's Apple in this case.

Speaker B: Once you standardize and streamline a couple of things, life just makes it a lot simpler at the end of the day. One thing I am particularly curious because not many of the people on this show have actually had this opportunity where they're, let's say they're. Was it both your parents who were verse in dos?

Speaker A: No, it was, it was even, even hardcore. My mother would uh, my mother would give my father the card to go get money from the atm. Uh, like that, that level of tech outsourcing. But since you know, they're not old school, my mother's a full time doctor. Like they're both like really same level in terms of uh, you know, responsibility for life. But they kind of have this deal with who is taking care of what. And my mother like 100% outsourced every tech decision to my father and also every financial ones, which, which now seems very classic but you know, that is, there's, there's like this unspoken deal like okay, if there's uh, if there's a concurrency or attack involved, it's, it's him doing it. So he's also an. And he obviously, he likes his role like oh yeah, now something to be tinkered with or something to be organized or, or with with the numbers. So. Yeah, cool. But at the same time she was always in charge of all the logistics like, which is not easy for both with three kids. Um, so yeah, it's funny like they had their deal, but no, my, my, my mother was still she. The whole setup should be so that she can, you know, just live in it. And then for example, the team in the cabinet needed to, you know, she just enters, barks this and that and then you know, the team needs to know which I mean imagine, yeah that they were, there were MTAs and they had to, you know, dog like DOS commands together. Super funny. In the end they were the superstars in the sort of when other cabinets took them over at one point, because they were like so tech first. They didn't, I mean they didn't have a mouse at first. So. Everything on Terminal until the last year is super brilliant, people. You know, it's a bit like these clicky coders versus Emacs and Vim coders. Uh, that's like, the efficiency is kind of impressive.

Speaker B: Yeah, they didn't learn it from school that they. Or, or like the way I was putting is more that like you have something that I guess you kind of require for your business, you require for your needs and then you kind of use it and then you optimize knowing the product and then you eventually are quite verse. I've used the word verse before and the idea of me using that is mostly just being efficient and familiar with, um.

Speaker A: Yeah, it's, it's, it's, it's a bit like uh, well, what's the best tool and the best way to get there? It's like, this is funny. I found it super analog. But I, there was only one recent story that I heard from the same uh, direction. It was a friend of mine and he, when he started working in investment banking, he was sitting there and then, you know, his boss came unplugged the mouse and I was like, okay, all right, there you go. And he's like, how can it work? And he's like, Axle has so many hotkeys. You can use your phone to look at them up. But be quick because from now on we expect the speed that we expect from someone working only with hotkeys. He said it was the most brutal week in his life, but it saved him like many other months in life later. Um, but uh, yeah, that's learning on the job and learning to do it quick and I kind of like it. It's uh, it's brilliant. I, I always love them tooling. I'm not obsessed with tooling anymore. Before it's like, you know, okay, which tools do we use? And the problem is, I mean you can actually go into an over planning rabbit hole there, like comparing the tools all the time. But at the same time I'm still, you know, I'm, I'm having my own setup here. I look at like my, my routers, my multi one, my, you know, this, this. I, I like when the setup works. So. And I have like my own email chain tooling and whatnot. It's cool, it's cool when it works and, but then I normally don't want to touch it ever.

Speaker B: I can imagine that it's just everybody Gets a rabbit hole of like, oh, here's my special command keys, here's my special hotline. But dot files have to be installed this way. Like it needs to be on a new, brand new machine. Like, trust me, we've all gone through that little rabbit hole a little bit. But um, but no, there's so much to cover. One thing I, I uh, always love, it's not really a tangent. Anything is that like we do code today. We do code over the years that we've been doing code, everybody remembers their first for Loop. I guess. Um, what, what was, what was the language you've um, you've used to write? I guess your first for Loop for me was I'm a very boring person at the end of day. I drop a little bit of like my own history is that um, I've never grown up really like coding. I'm a late bloomer. Um, my first official coding grade nine, around maybe 14, 15. That's. That was the first time we saw HTML, so that was pretty cool. But then nothing until university and university was the first time I worked. I wrote my first for Loop in university, which is quite late. Um, we were a Java school. So basically that long story was to ask you like, what was the first language you've written a for Loop with?

Speaker A: Uh, I, I'm quite sure it's basic. Yeah, I'm quite sure it's basic. It was Basic. Yeah, it was Basic on, on dos then, then quite, quite early Turbo, Pascal and I, I did top scale for quite a while. I loved it. Um, uh, we once actually programmed a little virus and rolled it out in school to everyone that didn't like us. Uh, like it was a bit like the nerds revenge. But it had to be at one point. Like, you know, people growing up, people are like physically superior and, and bragging and whatnot and you're like, oh, your computer doesn't work. That's, that's super sad. Let's. Can I, can I help you with all. Yeah. And our antivirus tool basically just deactivated our virus. It was brilliant. It was just brilliant. It was this little revenge we needed. But that was Turo Pascal. There was to Pascal phase and then um, quite early php, I have to say. I mean, you know HTML, if you grew up with BASIC and Topaz card and people say I program HTML, you're like, yes, you are marking up HTML. It even says markup language. So it's not coding. But that being said, I always had a lot of respect for um, people coding HTML because Front end drives me nuts. I mean the good thing is you see what you're working on but at the same time. Oh, uh, damn. When does this button finally move? I always have the feeling that. Which is not, not, you know, it's not, it's wrong and it's not merited in any way. But I always have this feeling like front end is not value adding. Like you know, we built the machine now, now, you know, paint it and. But this is obviously absolutely wrong. But you cannot get this, this, this random feeling out of my head. So I'm like front end is tough but I respect people who do it. So PHP was the thing because um, I mean I started my first company at 18 so 22 years ago. Um, with my neighbor is super random. Was also my parents lawyer. Not that my parents needed a lawyer all the time but it was a friend and he was like the lawyer friend, neighbor coming over and he wanted to do something on the Internet and he was literally like uh, you know this, oh, uh, I'm a lawyer and I want to do some stuff on the Internet. How do I do it? And So I asked Mr. Hacker in the neighborhood and we started this company when I was 18 and back then well, there were other stacks but I kind of still, I still respect the Facebook stack. I don't know if you remember like for a long time it was literally LAMP Stack, Linux, Apache, MySQL PHP to the moon until they had like half a billion users. And I kind of love this stuff. I love it like when people use existing technologies because they just work. It doesn't make for a fancy tech deck. Every investor would say what? But it's the stuff that actually scales. A funny side story. We're still running this company I'm talking about. We. It actually processed half a million Volkswagen Diesel scandal claims and it's still MySQL PHP. I'm just porting it right now to PHP8. Oh shit. Now I opened my door to some hacking. Uh, tries but no, um, no, no, it's, it's. I love, I love working technology. Why go fancy? So yeah, and then I start to PHP for quite a long time. I mean it's a bit random because it's a hypertext preprocessor as the name says. But even like my home stuff, my. I need to calculate this and that. I need to, to convert A to B. I always did in php, which is absolutely bonkers. It's a bit like you know, taking your car to go to toilet. But uh, but um, at the Same time. Yeah, exactly. It works and it has so much, you know, built in and the libraries and stuff. So that was like, uh, I found it super practical and I mean seldomly the problem was really that you needed more speed than you just, you know, go a bigger instance, who cares? Or something like that. Um, yeah, but yeah, and then other stuff included more like, you know, point, point, like point wise gigs. I once did my own router firmware. There was obviously C, but I would never say I'm a C C coder. Um, and then JavaScript obviously a lot. Um, and uh, even though it's not my style, I mean more like old school, plain vanilla JavaScript and all the rest. I never really got my way into it. And now, now comes the big one and then, then I finished that. Um, I still hate objects. I still hate objects. And you know, now that's the point when I totally accept that people say I'm an architect and software engineer, but no, not a coder. I get it, I get it. Shame on me. And I, I love algorithmic code. I love procedural code, functional code, the object stuff. Uh, I don't know, I mean I get it. And I now sometimes when people say, when I come up and say, hey, let's make an object out of it or a data class, it's like, oh, wow, we must be, we must be really at the limit. But you, I don't know if you remember, it was the time when you had perfectly procedural code, but someone had to break it up and make objects out of it. So like 50 people could work on 800 objects. That could be like one meter of script. And so I was a bit like counterculture there. And until this day I always write the algorithmic parts and the, it was quite like it. Like, you know, there's the chicken input, put, put output, uh, and then, uh, and that's, that's the stuff that I like, you know, APIs. This is in, this is, this is tak, tak tak. And then this is out. And not like objects most of the time are too messy. So this is why I never went into, you know, coding career, so to say. But more like build systems.

Speaker B: No, that definitely helps, uh, understand a lot of it. Like especially I definitely know all the different clans. The clans that love the oop, right, the object oriented programming. There's an absolute clan behind that. They love that kind of stuff.

Speaker A: Oh, I love, I love, I always call it, I always call it the, the ghetto type languages like where you can do whatever you want. That's what I loved. In php, A, A equals 1, B equals hello. C equals A plus B, it's 1. Hello.

Speaker B: Brilliant.

Speaker A: Exactly. My humor. This is what I love. But funny enough, my wife is in the exact opposite camp. So you can imagine how often we, you know, banter here even at home with this. Uh, and I increasingly get it. So I'm, I'm, you know, up there, I'm not religious, but if you ask me what I prefer, it's like, ah, hell, they. Yeah, you know, she's like, alex, did you include typecasts again? Yeah, type cast are exactly the bridge between your world and my world. Um, she's like, no, typecast the haram. Okay, okay, sorry. So now what is this? Oh, float. But it's round. Yeah, uh, just put float. Is this. I love it. It's, you know, this is, it's, it's the thing that makes you. This makes it lovable. You know what. But you know, funny, funny side note, you know where I got to love the SMS because I, I was like laughing my ass about decimals and then I founded a fintech and once it's about money and the decimal too is actually the only thing that doesn't fuck it up down the road. If you calculate monies, that was the interesting part where I, but I finally got the reason for decimals because if it's about money, then you actually need less than float or double. But you cannot just do it like calculate everything in cents. So decimal is actually brilliant. And finally it has a purpose. Well, and you know, finance and banking is not like the smallest sector on uh, earth. So. Yeah, bing. Now I get it.

Speaker B: It's very like, interesting to get the, the behind the scenes right when you're mentioning php, when you're talking about basic, when you're talking about all these languages kind of help you build, I guess, like not only your character in one, in one sense your identity. And also tangent, because this show is all about Tangent. People underestimate PHPs like presence in the Internet. As in like WordPress for the people who love diving into these stories. Like WordPress nearly dominates till today the Internet. And that's fully php as well. And just like all your Facebook and everything. Um, I think Facebook nowadays they have their own wrapper around php called.

Speaker A: Yeah, it's V8. They have a php to c live compiler.

Speaker B: Yeah, uh, yeah, so that's like, that's why they're still running with it. Surprisingly so. If there's anything that I know is PHP is not dead. People love to say this joke all the Time because obviously tech people are very funny.

Speaker A: They say, you know, the, the biggest oximeron is a good PHP coda. Yeah.

Speaker B: What I do want to call out is one thing that was very cool is that we did have a little glimpse already because so today you're, I'll use the word maybe serial entrepreneur. You may not use the word, you may use the word, but from my understanding is that a lot of it is on LinkedIn.

Speaker A: I use it, but I have to, you know, LinkedIn is a different art form. You know, I, I, I would, you know, if, if my parents would show me a profile like this on LinkedIn, I would cringe. But that's the way, you know, if you don't cringe, you have a problem with your LinkedIn profile. It should be cringy.

Speaker B: But the thing is like there's already glimpse in terms of, as you were saying, just working with their neighbor into building uh, a tech that we're going to talk about in just a second is that a lot of this came from, I guess like being a tinkerer growing up, playing with all these tools. You're talking about Legos. You're even remembering very clearly your first computer. This is one of the favorite thing I like to talk to people about. Um, I was, we were a Windows household and today of course, uh, I use Mac OS way too much, if anything. So one thing that might be useful is that um, my boring story is comps go into software engineering and boom. Till today I'm still doing software engineering. You may or may not have taken the similar path. So maybe just a quick glimpse in terms of was it Compsite you got into May? I'm gonna guess the answer is no. But what brought you from I guess like tinkerer as a child in your neighborhood playing around into education? Was it, Emphasize it. What did you get get into as a, as a major when you got there?

Speaker A: Yeah. So, uh, well, one has to know that back then in Germany we still had the diploma system. So it's a bit like bachelor master integrated. So you couldn't say I do my bachelor in that and uh, my master in that. And there was also not a major or minor. I mean you could obviously combine, but you basically did a two diploma in medicine, so to say. Yeah, and then that's, that was like masters but everything included and batteries. Um, so back then it was funny. I mean I, I, I, I, you know, thought about doing Comsci, but my mother, arrogant as she is and lovely, she's like, uh, she's so much into. She outsourced tech so much that she was like comsci. Com sci isn't. What did she say? Auxiliary science. Like literally it's an, it's an auxiliary science. So to do something, you know, um, not do something real. She, she wasn't looking down on it, but she's like, no, no, learn a primary science and not an auxiliary. Like super. It's. It's old school thinking, but it, it makes you think a bit but. Because also what she always preached is like do something where you're not perfect at, but where you learn a lot. And there's one thing, I mean, obviously if I now say engineering, you will laugh because everyone thinks I'm. I'm good at engineering, but I was like good in tinkering in some parts. And also very funny. This is super random. Um, the only, the only part of school, like back back then where I needed. How do you call it, like, um, help. Like home school, like a home tutor was maths. Actually math was just like. I mean fast forward. I now basically have a PhD in applied math. But um, uh, math was, was the problem. So she said like, hey, you know, go for, go for what you don't do. Like as a walk in the park. And there was other stuff. You know, I, I went to a very old school with Latin and ancient Greek. So I have uh, as part of my degree is literally Latin and Ancient Greek. Which is brilliant because until this day every technical word for me sounds like German. It's like if it's hydrophilic or lipophobic or whatnot is. It's actually super cool. I love it. Not to, not to pose, but no, when you read papers on everything could be a linguistic, anagrams, whatever, oxymorons, it's like, yeah, cool, I know what it means. Um. And um. No, but the thing is that that was all quite easy and. And math was like for a random reason. I think it was because school math is not math, but calculus. But that's a different story. And um. So yeah, I enrolled for um. It's called industrial engineering, which is. They always make this joke. Is 80% engineering and 70% uh. Um. MBA. And uh, how do you. MBA, economics. Um. The funny thing is it's actually not untrue. We compared it later. Like just the amount of lectures and tests we had was actually that. I mean it was a very famous thing. It was the kit. It's like, you know, that is for a reason. They, they, they kind of say themselves as a German mit and they are quite well known for many things. I think the first computer was there and stuff like this. So it's actually, it's a serious institution. Um, and um. And yeah, it was a bloodbath like this. You start with 600 people after two semesters and 150 are left and so on. So the beginning was super brutal. It was like this absolute high pressure, high octane kind of university. And obviously I failed in math at first. It was brilliant. It was like, holy. And the thing is then I had like my old friend Steve, who's, who's out now also quite a, quite a well known founder. And he came to me like, okay, Alex, you know, we have to stop this. You always explain to half of us how this stuff works and then you up the exam. Let me guess, your problem is, have you ever revised in life? And I was like, revive what? He's like, yeah, okay, there we have it. And he told me, you know, the first rule of basics, basic physics performance is um, work divided by time. So you know, you're just too slow because you don't have to practice. And that's. And not literally. And then we started, you know, revising, going to the library and training, training for the exam. And then I finally got better and everything works. But it was, I mean, shame on me, you know, I was so arrogant. I never revised my me for that. You, uh, know, I thought like when I understand something, it's okay, I understand it. Got it next. And then obviously in the exam you totally fail it because it's more about being trained to do this on speed. Yeah, well, um, the cool thing is after this, you know, after going through, you know, the storm at the beginning, um, it became very open. You could basically mix and match from 500 different lectures what you wanted to do. And that's pretty cool. So some people, for example, they did a color chem, color chemistry and um, art history. And they are now at Sotheby's, you know, in the restoration business. Some others did, I mean most did everything I need for consulting. So it's, it's well known that many people from there go to McKinsey, but they're there and the other consultancies, but they're the guys who at least know a bit of their way around tech. So that's, it's good that uh, you know, we have this. But for me it was energy, energy, energy, energy. I even my um, my business part in the studies, I put like, I think energy business. So I basically only ever thought about energy all the Time like from solar cells to nuclear power plants, coal plants, whatever. I wanted to understand all of them down to every component. And in the end I did. I literally we had like this four hour exam where we needed to, to draw like the components of power plants and calculate them. And it was, it was pretty cool. I, I loved everything about it. Problem is, um, and, and you know, remember all the while, um, I was running this, the small legal tech back then, it was still very small with my lawyer neighbor. Small side note, we were so early that uh, yeah, cool, let's reach out to uh, the German lawyers. How many do have email addresses? Yeah, 10. So the business started very, very small. But then it kind of, you know, we grew with the market, but we had many, many years. So end of my studies became more interesting. But on like during that I just give, you know, remember I still have this business at the side. But then what do you do after your, when you graduate? For me it was clear, um, or at first I thought, okay, um, if you know everything about energy, okay, where's the job application for an energy visionary? Well, there is no, you can, uh, you can work as, you know, the project engineer for a company that produces the screws for this part of a power plant. But yeah, that's not exactly the big lever to change the world. Um, but I realized now I, I'm also very curious guy and I want to go further. And as you see I love explaining stuff and, and I can talk. So I, back then I, okay, let's go into academia. I wanted to become a professor. Problem is we're talking 2008. You remember what 2008 happened? The super crash that made like all of the first years, uh, in consulting realized, ah, you know what, let's hibernate in academia. So my professors taught me, told me normally they had between 10 and 15 applications. Now they had like 3,000 and cannot even sift through them. It was absolute bonkers. So the problem was that many of the PhD program applications were down because everyone, like at least in Germany, in Germany, I, I, this is, I mean, I don't know, maybe it was the same in the US but it's a bit sad that academia is more like an opportunistic backup instead of, uh, you know, I want to teach, it's more like, oh, let's, let's hibernate there. Uh, winter is coming, let's go to academia and then, then go back again. Um, and with a doctor on top. So yeah, cool. But there's people who really want it, um, but yeah, finally I found, uh, a very good spot in the German Aerospace center here in Berlin. That's when I moved from classroom to Berlin. Then, um, during Kaso, I had been in Reunion and Chile. That was like my, uh. It was not integrated. Organize it yourself. And back then we had time, you know, I was one year in Africa and half a year in Chile. It was pretty cool. So I learned the languages and I, I could really connect. And. And I always went back to both. And um. Yeah, but the thing is, uh, so finally I found it and it was, you know, same problem as energy, if not bigger in transport. So I found myself in the Institute of Transport Research where it. My, uh, contact with AI started. We trained. I tried to train a neural network to predict. Back then the project was predicting who drives what car. Because remember 2010, we were about to shift to electric cars. And everyone was curious to know, okay, what are the incentives? When do people switch? Range, Range anxiety, and so on and so on and so on. And this predictive model was my product still in use. That, that was. That was me. That was the thing. And, uh, I tried to do it with neural networks, but, man, you could. They were very good CAC picture detectors. And we had a lot of fun detecting cats. But they weren't exactly good for predicting how, uh, ad or how 40 million households will behave in terms of car buying that we did with, I would say, the conventional tools. And then, um. Well, the rest, uh, is history. 2013, I finished my PhD, which was big data econometrics for. For two funny reasons. The one is like, okay, my. My supervisor at the Aerospace center, he basically had a professorship for economics. So I was, you know, it's. It was easier. But also, I have to say, I. I kind of voluntarily switched to economics because at one point I was really disappointed with engineering, at least in terms of the academic way. You know, it's funny, in engineering, I saw a formula once that is like, um, this and that to the power of eight. And it was. I think it was about how helicopter blades, um, you know, what they do with the air or whatnot. And I said, wait, wait, guys, what the fuck? There's no process on earth that is higher than to the power of four, which is, uh, radiation. Like, you know, it's super funny. There's stuff, you know, that is quadratic, there's cubic. And then like, power of four is the highest you can have in a natural process. It's radiation. Why the. Is there to the power of eight? Yeah, because it fits the pattern Best. So, okay, you don't understand it. You just do. Curve fitting. Yes. And for me, that was like engineering. Um, okay. It's like. No, no, I, I'm still an engineer. I still love to think. I still, you know, did you power it off and on? And by the way, that works for everything. That even works for servers. You know, we power it off and on each night and it drives me nuts. But that's okay. But that's not academic anymore. You know what I mean? It's like there's an academic word and then there's the real world. And in the real world, things were. Things happen and, and everything is chaos. It's okay. But academic means understanding the world and telling me that the process is something that is basically just a curve fitting that is not understanding the world. Whereas in economics, you would never get through with this. In economics, you have to, you know, this is why you see the logarithm so often. Like, oh, the process is ln of this and, and this times this and, you know, then it barely fits. But at least the theory holds. This is why I said, okay, my mathematical model will be presented not on the engineering umbrella, but under the economics umbrella, where the stakes are and where the requirements are super high. I mean, I, uh, had 85 factors hypothesis and, and everything had to be proven and not like, yeah, put, put a Power 8 and that your curve fits. So, yeah, it was a long story. Why. Why I have this strange combination of my engineering degree and then this economics, uh, PhD. But, uh, I'm, Yeah, I'm very proud of the combination. And it's me. It's perfect. It's exactly reflecting who I, um, am.

Speaker B: The fun part is you didn't mention is like even just trying to balance all of it as we're seeing, like, so then like eventually going into your PhD at the institute, it's, uh, not even the institute, the German Aerospace Center. Sorry, that is already fascinating to begin with. On top of that. And then you've dropped this one. Maybe there's, there's. There might be a story in parallel all this. So while you were studying, you were also building this fun.

Speaker A: Fun.

Speaker B: I don't know if you would describe it as fun, but this product that is about, uh, German law and then there's going to be a lot of tangents maybe. I mean, maybe it has something to do with the German German aerospace. You also mentioned Volkswagen previously as well. Um, what's this? What's this kind of like crossover? So on one side you were studying PhD at the German aerospace, um, center. And also with your neighbor. The origin of that story is that, hey, they were a lawyer. And then there was maybe services that lawyer needed in Germany, but maybe also around the world, to be honest. So were you building websites? Were you building, I guess, like just software. Were you building just like Excel sheets back then? Like, what was a bit of that highlight during that period? Because it's very interesting.

Speaker A: That's a very funny story. And you will see how they, how they. I mean, honestly, uh, it will interweave in a way that I couldn't have dreamed of. Um, so. So the thing is, this lawyer, he wanted, his idea originally was, um, okay, let's build a directory of lawyers online. I mean, we're talking 2,000. Yeah. So really? And I said, ah, there's already, there's already three. I don't, I don't, you know, build copycats. Stupid. The first nine all got rich. But, uh, you know, I was, I am arrogant by design. So, nah, it's already there. I don't bid it. Uh, but the funny thing is when we wanted to meet for that, he was late. And, uh, here I played and I asked him, okay, no problem, but why are you late? Yeah, I had to go to court and do this thing. Super stupid. Every lawyer could do it and so on. I was like, okay, so every lawyer could do it. Like, literally. Could someone have replaced you? Yes. Okay. Why didn't you have a replacement? Yeah, because I, uh, don't. It's tough to find. You have to phone everyone. And I'm like, I have an idea what we will build. Um, not really.

Speaker B: Really.

Speaker A: We invented the one lawyer replaces another process in Germany. And actually it's kind of a huge thing because if you think of it, for example, I sue Amazon here on Amazon. The German, uh, headquarter is in bad hairs, where it's basically Frankfurt. So that means for my, I don't know, 50 bucks bottle here, that is not optimal. My lawyer drives down there, holds up the hand that basically restate what's already, you know, in the dossier and then drives back up. Lot, uh, of cost risk for me. A lot of pointless emissions burned. Doesn't, uh, make sense. Now if in the old process before us, uh, the lawyer call, like, goes through his, his personal phone book. Someone in Frankfurt. And. Yeah, okay, do you have time? So just the inefficiency of finding someone. But then this guy would say, okay, then I'm, uh, also council of record, right? So he gets, like, down the road, half of everything, at least. Even though he just once held up his hand and we said okay, we can do this better. And Jurgen, my partner, he um, did the legal engineering, so to say. Like okay, which contracts do we need and how do we organize that while I build the whole system. And then at first no one was really interested, like two or three some. But then it became a bit of a thing like to you know, um, do uh, all this in Germany. It's called like the Americanization of law. Like the access to law for small people. Like before it was like lawyers were to for disputes between more well off people and whatever. I remember when I was a student like oh shit, the guy comes with a lawyer. Like uh, danger, danger. Now it's like hit a website and get the money back. Um, uh. And uh, for all these services like flight ride. Yeah. Or reclaim this and that, your landlord overcharges you. What? Not what not what. That's all us in the back end. Because whenever these services, you know, need a lawyer actually going to court, I mean, you know, how do they do it? And then, and then came the Volkswagen scandal and the rest is history. Like Volkswagen, you know, we had like totally uh, in Germany with like millions of lawsuits and half a million of them were actually routed through our system. Because who else could give you a lawyer in Wolfsburg tomorrow to state the claim? It was brilliant. It was super fun. Now comes the crazy part. By then I was working at the aerospace center and my predecessor on exactly my post in the Institute of Transport Research was the guy uncovering the diesel scandal. And I didn't know. So imagine all of a sudden it hits me like, hey Alex, you man, like the red line in your career is not for real. Like uh, you knew all along and whatnot. And it got even better when then after the aerospace center, I found it fair. Um, like the fintech and we took, took the insurance companies head on with the pension products. And then later, I mean some people said on stage, no, I don't sit with this guy. I don't sit with this guy. He, I don't know what's. What was wrong with him. Like maybe an insurance company, you know, screwed his parents or anything. But this guy from the age of 18, he's following his Mephisto master plan to totally insurance companies. Because obviously who paid for the details get in the end the insurance companies. And, and I'm like, oh wow. Ever since that I believe in chaos theory so deeply because I didn't realize that you could find such a red thread in my life. This is so brilliant. I really didn't know it. And in the end, I mean, when I. When I look at it, I'm like, yeah, that really must look super bad from the outside. As if I have planned it. This is absolutely cool. But then also, you know, sometimes, Sometimes now, I use it, like, for when you have tough negotiations, like, yeah, you know, that I'm the guy who. Okay.

Speaker B: Oh, yeah, sure. Use as leverage, if anything. Exactly. The other fun part is that, like, this legal environment is just very different than what you were studying or researching at that time. Like, you know what I mean? It was just like a completely different, like, dual. Duality of your life at that point. You're like, oh, let's handle that as well.

Speaker A: I call it like orthogonal stimulation. I. I love the orthogonal stimulation. I. I was even like a hobby doctor at the time, like, always, you know, like, you know, doctor appointments are hard to get. And so I, I always did, like. I mean, never, like, super serious stuff. I never stitched anyone up or anything, but, like, the small stuff. And then I. In the back, I think, with my parents, like, okay, this is what I would diagnose, and this is what I prescribe. And then they sent me over the prescription. Okay, this. Cut this out. But no but, but the thing is, um, I love, like, whenever you skip or connect disciplines, there's something in it. There's something in it that, uh, that, that, that. That stimulates you in a way that's cool. And always also, my wife was like, how many projects are you doing at the same time? And, you know, I'm, I'm. I'm okay, stress resistant, but I'm also pretty lazy. This is why I automate so much. But, um, I like to have, like, you know, rather I have, like, impulses from 10 different points and then only do the important stuff. So I'm a focus guy as hell. I mean, there's unpaid m. Unimportant unpaid bills here. There's a whole stack. But I will pay them at one point when they escalate, because I cannot care about this now. I care about this other project from this other discipline, uh, that I, That I find fascinating. And I think it's good for my brain. To be honest. I realized that the more I have up to a certain limit, the more I have, the better I am in all of the things. There's only one thing that makes me crack, and this is like, organizational overhead. Like, but now we have to organize the travel. That kills me. But I could found another company in the meantime. No problem. Um, it's like yeah, this is why um, my wife really helps me a great deal with this stuff. But always like I try to get rid of day to day organization. Paying your taxes, horrible. Getting your card for, for winter ready. It's for me like it's the task of the year.

Speaker B: It's like there's nothing really in between. It's either like, either like there's very simple things like oh, make a coffee, that's totally fine and then the next thing is like found a company but in the middle between all of that there's nothing, nothing like there's nothing that confused on that. One thing I do want to skim over before we go into the origin story of Pomo AI though is that even after, I mean you are still doing a lot of your lawyer related project, if anything, um, that is still ongoing. That's the passion that I could clearly hear through the mic. If anything that's something you love diving into. Um, the between that and then you did mention you did work on a pension related slash fintech related product. We could skim over that a little bit just because that does proceed into Puma AI just a little bit. So one thing I didn't notice a lot from what you were doing is that there's a lot of like commonalities and patterns of how you work. Whether it's studying for a PhD, researching for a PhD, whether it's working in the law industry, the law hemisphere, if anything. And then going into this like pension fintech world, you grabbed a lot of like patterns and framework I'll bring into it. So when you got into that, what, what was the entry point?

Speaker A: Um, 2013 I finished my PhD and uh, I wanted to become a professor. And I negotiated and we were close. There were some ideas. My problem was that I still believe in. They say it's the ideal of Alexander von Humboldt, my namesake, um, who said um, you always need to combine research and teaching, then it needs to come from the same mind. And I deeply believe that and I love that. So I wanted to do first class research but I also wanted to teach. And back at the time it was this trend to have like industry projects all over the place. Like you know, get uh, more funding from industry or apply and it was less like okay, foundational funding and let's just develop some cool stuff. But it was more like um, Hustle has um, a BMW, so you do project together and then you only do the industry project and who cares about the students. And um, my one problem was that I already had like a Potential lecture. But since it wouldn't count for work, I would have to take vacation to actually do my lecture, which even algebraically didn't work out. Um, and I realized, okay, you know, maybe. And also, let's say in, in this discipline, I mean an economy is a bit better than in engineering, but still, um, you could use more women there. And uh, you know, I'd say with affirmative action, my chances weren't the highest. But also, honestly, I would, I would happily, you know, yield uh, my place to, to female professor because I think it makes sense to kind of, you know, uh, equal the scales there. And then my supervisor, my doctoral supervisor, he said, you know what, Alex? Uh, I will give you my whole nest egg here. Here's some money. Go fund a company. You're best in that. It was super, it was super crazy. Like the show of, of. Of. Of conviction. Uh, and uh, because back then, um, Jens approached me. I, I didn't know him. Funny enough. He is the. He was studying law in Hamburg with the son of that lawyer I had the company with. Yeah. So, so really like super random. And, and when Jens said, hey, I'm. I want to found. I want to do an Internet company, then Alex, like the guy at my age, the son of the lawyer said oh yeah, yeah, if you do want to, want to do something with Internet, you need to get to know Alex. Like, you know, go to Mr. Internet. Super funny. Um, and uh, yeah, and Jens called me and he came to Berlin and he pitched me this idea like, hey, I want to disrupt the German pension system. And I was like, literally, wait a second, I get a new beer. There was literally the legend of like uh, his pitch in my first answer. Um, and I realized it's the most boring thing I've ever, ever heard as a pitch. And then I got home. Yeah, no, I mean the end is super, but, but, but the, the. The idea of let's build. Let. Let's build pension plans. Like holy. And then I got home and in this. Yeah, and exactly that, uh, you know that that stack I meant. I found like letters from the aerospace center with my pension plan and I'd never opened them because. Or, or one I had opened the first and then don't understand too boring away. And I realized that I had a pension product unattended that I could have done more if I had attended to it. If I had understand it, understood it and produce something with it. I was like, holy fuck. So the second largest tickets for German household. Well, obviously after housing and sadly the car. We are Still Germans, um, is pension and I mean for housing and for cars there are markets, there's competition. Yeah, there are some problems, let's not delve or dive into that. But I mean at least it's like more or less transparent markets. Whereas for pensions, at least in Germany, it's absolute intransparent. There's some random broker just selling you some stuff and it's madness how much money gets lost in, you know, the market inefficiencies of not having honest competition and good products and good prices. Yeah. And so I, and once I, you know, once I smell blood, I smell blood, then it, it. No, I mean originally the idea was to help Jens. I mean who is ah, a super, is very clever and unprofessional, lovely guy but he's really a non tech and he admits it. So I got disclaimer, Jens is now my cfo. So all good. And uh, we still love each other. But um, he, I mean at first he gave me, he gave me his plug, his power plug. So I plug in his laptop. I said that's too much, that's not what a CTO does. But you get the idea. I wouldn't have put it past him if he actually needed me for that. So yeah, the first idea was okay, uh, uh, me not Internet, you Internet, you do Internet for me. But um, uh, then it quickly started that I wanted to, you know, also help with the product development in terms of financial product. And once you send me down the path it escalated completely. So we developed a whole new product not only in terms of actually, you know, the tech stack wasn't like I say again, Facebook stack. Really? We did this with php. Why not? It just works. Um, and then, and then you know the, back then the VCs were all about oh there's like um, uh, you know, how's it called? Like the, the um, um it was the time when you had like all these cleaning startups like um, this and that mates and so on. Um, and they had like their, I got some decks from VCs and their tech stack was like hey guys, is that, is that a Java lab or is it like a cleaning platform? Why do you go overboard so much? Um, and we were like yeah, php. Like the tech stack was like a simple page but we said let's use the resources to really dive into this financial product. So I, I, we built the product, we had an insurance company, we had a bank. Like um, we combined them, it was absolute bonkers and then we need to calculate it. So I, I wrote My own calculation course for the product. So we had the first retirement product. We just move sliders and see in the live was. It was a modified RALJS library I, I built for that so you could live see you know how your pension product was today. This is not a big deal but we're talking 2013 where people normally would get sent, you know Here is your 50 pages that you don't understand and instead they move you know a slider and see you know their pension involvement. And I always like to say I had once built engines at Mercedes. We dropped this part but it was, it was a longer but deeper kind of internship. Um, uh, during the, during studies and um, where by the way again not shitting you, I was uh, working on the exhaust gas management and we knew that down there the, the, the passenger car guys were faking. So like again you know the red thread and stuff.

Speaker B: But it's always there.

Speaker A: Yeah, I can get rid of it. But um, have I always said okay, building an engine or calculating physics or trying to calculate the German tax code is actually pretty close to each other. Physics was easy I have to say. Um, but uh, in the end we were like the, the pension nerds. I mean up, up to this day I'm still in commissions and in uh, on, on panels for pension product design because we really like you know we went down the rabbit hole until we hit the wall. So we are there now on the pinnacle. We, we even co designed like for the next um, uh reform they will do. It's about to come. Finally. I fought 10 years for it and now it's about to come that Germany, Germany. Germany's pension uh system will be reformed. We finally will see the light. It's no, it's really, it's. It's a shame what happened here. It's an absolute shame. I don't know if you've read about it, but Germany wasted so much money. People earn much and even save much and 10 years later they have half the money of Americans because the money goes into random broker commissions and the wrong products and we finally bring them to the capital markets again. But yeah, long story short, we built this company and found the perfect match for um, there's a Fintech uh, unicorn and we knew the founders and they're super cool. We are close to each other and in 2019, so six years after fund after founding, we sold the company to them and integrated as their pension division. They didn't have a pension division. They were lacking some features in the investment products and we sold to them. And uh, there was like really a nice, nice, cool exit story and we were all of a sudden part of a super big company. I was cool at the beginning because we were the star state and you know, the PMI was nice. Jens at one point dropped out earlier because obviously corporate stuff, finance law is more redundant than tech platforms. I stood on for the tech pmi. I still am like a part, part, part time consultant over there to help with the remains. Um, but yeah, you know, you're all of a sudden your tribe lead pension, uh, head of pension products company has 900 people and you have like 82. Cool. But I'm not a corporate manager so it started nagging that um, I need to do something else. And uh, so yeah, but then you

Speaker B: have the awareness to be like, okay, now what? Right? A whole head of division like the pension products, super important. But then your drive was to look a little bit deeper into something different at that point. And this is where Puma AI comes to be. I mean like a lot of what you ended up doing just builds up to where we are today with Puma AI. So Puma AI, this world of like LLMs, rags, big, um, data, I mean you can even throw that into there. There's a lot of ingestion, there's a lot of data pipeline, all that kind of stuff. Maybe before we jump into like the tech, techy, extremely deep dive, all the kind of coding stuff, what is the origin story of Puma AI? Not only just the name, but also what was the setting? As you were saying, you were just, did you just wake up and be like, boom, that's the idea? Or was there a little bit more to that?

Speaker A: No, uh, there is more, there's more. But first of all, thanks for um, coining that same term. I always like to also say, uh, it. There was a journalist once that asked me, yeah, but now you do something new and then you just want to sell it and I say, hey, sorry, I deeply despise this, uh, kind of negative, uh, perspective on it. Not on people who just, you know, like push up some bullshit to say that that's, that's totally fine. But I'm a creator and then there's administrators and I know that I'm a bad administrator. I'm m. This is fully fine. This is why I deeply respect good administrators. Um, but for me as a creator, I shouldn't, you know, just transition to administrator because now I got the merits or whatnot. No, no, I'm bad at it. So I need to, I mean, I Always want to hand over something good. And I always prepare my handoff perfectly, but then it's hands off to create the next thing. And obviously, uh, well, why create Puma? So now the story is I, I had some time on my hands. Like I, I had a little medical problem, um, had a perfect successor at Raisin. And I did the handover and I took some time off like, or mostly off, like, reduced my hours and started going back to. Okay, remember what I told you about Mere M PhD AI was a good cat picture detector back then. Maybe it can now do more. I had followed GPT2 a bit dabbled with GPT3. And then we said, okay, see what we can do with this. And um, it was especially, I mean, hallucinations were a thing before ChatGPT made them public. But everyone knew, okay, these things can hallucinate. Um, and the thing is, uh, our idea was, okay, how can we get rid of hallucinations and professionals know, training, Forget it. You don't get hallucinations away. CGPT 4.5, more training. There's kind of a digressive return on that. Fine tuning brings you skills and style in the model, but you don't get the hallucinations you way. There's only one way, and that is context. Okay, that being established, we looked into it because we said, you know What, I have 10,000 law firms in advocates. You know, we go back to square one. Uh, the company, we have 10,000, 10,000 law firms there. So if we just build a chatbot that does that, where you don't use your license, you don't get disbarred because of some hallucinations where we already have 10,000 customers. Let's try this game and um, let's first, obviously, you know, get a grips on it was not even chatgpt out there. So let's see, uh, uh, what we can build. That was 2022. And um, we started like, okay, like installing first systems, you know, like combining the models and stuff. And um, okay. And then we first realized, um, the problem is we cannot build this German legal chatbot because it's a content game. It's like, uh, especially in Germany. I don't know how it is in the US but in Germany only 3% of verdicts are online. Um, the laws are mostly online, but in a shitty format. You have to like, clean it up very much. Um, but especially whether know we don't have case law in Germany, but what we have is the codified law with a lot of secondary literature. So people say um, yeah, but according to this magnet, uh, that published this opinion in this book that, you know, you can only get when you have the subscription, but you should know it, otherwise you're in court. Um, uh, it's this and that. So, ah, ah, the law says something different, but it means this and that. And this is well guarded by a set of, let's say, five publishers. And in the end, it's like, it's a power game. Like, okay, who gets access to what. And it's funny and I have a name in the scene, but I don't know. Uh, it's not my style. And it's like, you know, only Titan negotiations. And I want to build tech, but I realized, okay, let's at least try to build something like that, because if we make it there, we can make it everywhere. So the idea was to build domain specific chatbots. That was the original idea. Let's say, um, we, we can build this for, for, uh, the law. And then, by the way, I would have sent my partners from advocates. I still have two partners. They could have haggled with the magnets that sit on the content. But I built the tech and I just give, you know, I have my first customer in Advocate. And not like advocate has 10,000 customers, but cool, outsource what you don't like. Um, and, uh, and then we would build like the, the medical AI that was also always a dream, by the way. It's prototype found my carpal tunnel syndrome. So, um, we also know it works. Um, but, um, yeah, domain specific eyes was the idea. And then now it's an important excursion, a small one. Um, every winter, I mean, if I let me turn my camera and you see, this is a rare exception of Berlin weather at this time. It's super beautiful. And normally it should be gray, gray, gray. Look it up. In the business book of world records, the three, the first three places on consecutive days without a ray of sunshine are Berlin last winter, Berlin the winter before, but in the winter before that. Um, so, uh, yeah, my doctors at one point said, okay, you cannot eat enough vitamin D to sustain. This here is antidepressants. I'm like, guys, guys, guys, before I eat antidepressants, maybe there's a more offline solution to that. And that is I just go to a sunnier place. Oh, yeah, interesting. And there is this little. This is tiny country in the Pyrenees. It's between Spain and France. It's called Andorra. Many Germans think it's an island of Africa, but it is, uh, a landlocked Municipality, it's a bit like, you could say Mount Monaco, but it's a lot simpler. So it's more like poor man Switzerland. Um, so a lot of, lot of ski slopes, a lot of duty free shopping, uh, but super nice culture. I, I like them, I really love it. The only country on earth speaking Catalan even, you know, Catalonia didn't make it there that far. But they speak Catalan and it's Sunny. They have 300 days of sunshine, uh, of which most are in winter. So every around Christmas, uh, or New Year's we move and we move literally for three, three to four months and we live there. So we ski a bit every day. We sit in the sun, it's nice. And we work from there. They have 10 times more broadband Internet than Germany, which is a shame. So whenever I'm in Andorra, people are like, uh, Alex, you're an adoro. How do you see it? Yeah, I see every pore in your face, the resolution. And it's like I'm in my mountain hut. It's like, ah, uh, so don't get. No, sorry, I don't want to rant. But Germany, if any decision maker hears it. Can you really mean that? Um, uh, yeah, but this is cool. So we are in Andorra and Andorra has one thing that is super fascinating. They have one platform, like accessible for everyone, with everything legal. So from the constitution down to parking regulation of each village and even the, you know, the speeding tickets that someone didn't pay. So everything's public, in your face. And yeah, but that makes for an absolute data trough. So if we wanted to build the first country where you can talk to your law and also, you know, since it's a small country, at one point you go to a bar and you get to know uh, the ex minister and so on and so on and they were like, hey, cool, if you make this, uh, hey, come on, we get some state funding and we are the first country where you can talk to the law. It's still a side project, but turned out not so easy because when, I mean, so we scraped 200,000 pages, we put them all into our rack, uh, that we had built for that and I think it was one of the better racks. And honestly like, you know, hybrid search and all that we already had like three years ago and you know, we had like versioning data governance, what not what not what not we kind of over engineered to be honest. But um, we wanted to make this one thing really work. And then was my birthday, my birthday, 6th of March, 2020. 4. When we said, um, okay, let's, uh, have a public demo. It was the most expensive birthday I ever had. And not talking about the drinks or anything. No, I'm talking about OpenAI's, uh, invoice the next day for the tokens. It burned. Every question burned more than a dollar. I mean, yeah, it was a bit more expensive than now, but, um. Problem was you couldn't find the right context to actually answer the question unless you included, like, all neighbors and all chunks between two chunks you found and whatnot. So you basically had to cast the net so wide that it was. You needed at least as many dolphins as tuna in your net, uh, to, uh, to, To. To. To. To get what you really wanted. And that's like, you know, I don't know that it didn't seem right, but we couldn't crack it. Um, what is important. There was one thing we forgot to establish. Wait a second. Um. Um, there was one thing I realized. We didn't even. Ah, yeah, yeah. When I. When I say we and when you say, how did you. How about this? How was that all possible? So, yeah, with raisin. I mean, I, I didn't have a crazy cash consideration in the exit. So we are raising shareholders now, um, which is cool. I always like to, you know, sit in the same boat. And I'm not actually good with cash. I would just, you know, invest it again in something, uh, or do a next project. Um, but it allowed me to, let's say, set up my private research lab. And I did that in Andorra. This is also the Andorra angle. It's not only this legal platform and me living there in winter, but also they have, like, you know, they, they. You can set up a company there and, and just work. I really like it. It's not like in any way a, uh, shade or anything, even if it's official. It's an official subsidiary of my German company, company. But, you know, the whole bookkeeping and, you know, it's just like input output. Okay, done. And if they get their 10%, they're happy. Super nice at this time. Thank you, Adora. Because setting up a company is a lot less hassle than, for example in Germany, where you are buried in paperwork. Still. Germany is still very, very complicated. And, you know, I just bought a fax machine to be able to, uh, deal with the authorities. I put it in the middle of the office so they really feel it when they visit. Um, but, uh, yeah, but it's, it's, you know, then this register and this. It's just too much overhead. And in Adora, it's like a lot less. So I did this Andorra Mountain research center, um, where then also I started. Well, I started paying my wife, or so to say. My wife quit her job. She's also brilliant coder. She's like, coded with long term experience. And I wanted her to join the project. And, uh, so she, uh, quit. So I sustained us both. And then another friend, uh, joined. So this is when. When I say we, I mean the first three people working on Puma. And it was already named Puma, the project originally. Yeah, it's funny, I always, um, I always wanted a company about something mathematical. I don't know, maybe like, you know, in your face. And in Andorra, they speak Catalan. And now it gets super arrogant. The word for apple in Catalan is poma. And. And puma also stands for power of maths. So the idea was this. This is why. This is where our logo also comes from. Like, you know, the apple with the square root coming out of it as the branch. And then power of maths, the AI was. Then later, you know, Jens said, you know, he's the business guy. Hey, you know, people don't get it. Power of math. All good and all fine and good, but you have to say something with AI. So even the idiots know that we do something with AI. And I'm like, okay, yeah, you win, you win. I. I would love the intellectual, like P. Power. Power of math. Yeah, it say A. Okay. Okay.

Speaker B: You get to pick whatever appeals to whatever target audience at the end of the day. So that actually does put it all together because, like, you were testing your GPTs, you were testing like, any kind of LLMs, and suddenly your bill is racking up. Uh, people see these crazy stories to be like, oh, wow, the bill is like $200,000. And it's like, where did that come from?

Speaker A: From?

Speaker B: And it's like, okay, that's a bit crazy. So eventually, with this kind of like, context, you're like, wait a second, like, this cannot be sustainable. There's no way somebody's willing to pay that much every single time for maybe just a demo that you were doing to some, I guess, remotely important people at that time in Endor. Um, and then this kind of brain power, you're starting to make these, like, connections in your brain to be like, wait, maybe you can make this more efficient. So now when we get to the point where it's like, we're going to talk about how what Puma does today compared to the original story of Being a chatbot to begin with or like multiple DOM main chat boss. That was like the original part. I'm gonna just throw some keywords out there and when even the word hallucination, if we talk about like a layman's

Speaker A: term for it, is like first with hallucinations. Just as a, as a connecting example. Um, I asked the original ChatGPT you remember like was based on GPT 3.5. It didn't have Internet access. So we're talking like many people don't really don't, don't remember it like the original original just like pure input output chatgpt. I uh, asked it who's Alexander Keem? And it got back with his serial fraudster from Germany and, and like what the fuck? So I thought like oh damn, is there really like I actually, you know, reflected on my own history. Then it's like okay, what did he do? And then it started something like yeah, there. And it was like half. But then who is co fraudsters. And then it made up something like super, um, well known German entrepreneurs. But half their last name was wrong. And so I could really feel the hallucination. Like it was um, it was strafing on the truth. But then I became a serial fraudster because I was a serial entrepreneur. It was just. That was like the. But. But yeah, no it really is. Don't get me wrong, I don't have any, any uh, skeletons. No. But the thing is remember when it says fraudster, even that it says it itself. Um means that later every token after that must you know, fit to that. This is why my co conspirators were best of first and last name of very famous entrepreneurs and very famous fraudsters. This was like the original hallucination. I saw it like holy fuck, that's brilliant. That's an art form. Ever since by the way, I'm dabbling a bit with. I call it giving the AI a stroke. So I actually make the. Make it go bonkers um for with some triggers. So it's funny, I still love this. The hallucination world is a bit like people on drugs or something. Um, so uh, it's a different thing. But yeah, hallucination, exactly that context I would say is everything beside the question that gets processed by the AI. Simple as that. Um, I think for, you know, when I explain context it is like when you upload your document while you ask a question to ChatGPT, you are giving it context. Simple as that. Like everything but the question. And in rag, um, one could argue that the AI gets the information like the AI goes retrieving. That's what we see with ChatGPT, like uh, searching the Internet for. Because I don't know myself. Um, that's absolutely that. Then it is, it is retrieving in a corporate rag. What you would do is you would say you have a database and you have your orchestration, your rag that you know, whenever someone with the chatbot asks some question about X, then your system would go to the database, retrieve the information on X and put it as context together with the question. So the gen can then answer. But yeah, I think we're very close there.

Speaker B: And the thing is like the more you talk about this, that really ties into what PUMA does today then. So even just going back to the part of like Puma, where does this sit? And of course one of the keyword is chunking that we're going to throw into the there. But maybe from like the basic understanding of, I mean the space is huge. What does PUMA AI do? How does that fit into this world at the moment?

Speaker A: Yeah. So also historically, let's, let's we can connect back to uh, the morning after that birthday, uh, birthday. Birthday. Cash burn. Um, and um, what we see, what we realized is okay, there were always either too many parts of the text included or not enough. But mostly we went for too many because we wanted to answer the question. But we realized, okay, how do we crack it? That we search for this information and the heading is up here, but then the table with the value is down there. And I mean, what do you do now? Do you include everything under that heading? Do we include everything left and right of everything we find? There is some problem here in how these, how the retrieval units are uh, structured. And there we have the magic word. So retrieval unit means what is one block? Like what is. Take Google. What is one search result? With Google it's easy. It's one URL. It's one URL and you get like one result is one URL. You can click on it. With AI, it's totally different. And not everyone knows this. It's not like I found it. Here's the document. Also the document itself is a bit huge. Imagine you upload to ChatGPT every, you know, you upload the whole Bible. Every time you ask a question it will actually stutter and say it's too large. Um, so ah, the funny thing is the, the search results or search units for rag or context engines or however we want to call it by the way, this is always a new name for it. Rag is so dead. Rag is so last Friday. Now it's a context engine. That's. It's cool. I'm cool with that. You know, I can now say rag is dead. The long live Rag. We just rename it, um, and it will be there in 100 years, I swear. Uh, but maybe by then it will have something name. But um, the funny thing is, so actually one unit must be one unit that you can embed. And this is the interesting part in the whole AI game, like embedding means you represent what is in the text in uh, let's say I always call it in brain coordinates. Yeah. So when, for example, I give you, let's say I give you Paris and then you get back the vector for Paris. And people understand vectors perfectly well. Vectors, by the way, is just coordinates. It's like XYZ and then, and then another 2000. But if I have to explain vectors, I always tell people, well, take the coordinates of Paris minus the coordinates of France, plus the coordinates of Italy, and you get the coordinates of, of Rome. This is how vectors work. And that makes super sense. This is also why people understand that you can have a, uh, uh, Spanish original text, but ask it English. Because obviously, as with this example, you can find in like literally information that, that correlates in the brain. And not only text by text, not full text anymore. The problem is that one of These embeddings, like one vector, can only be based on maximum 8,000 tokens. 8,192. But that's basically 10 pages. So if you think of it as one train of thought, 10 pages is a bit much. I mean, you have to be very verbose to put. I mean, I could maybe do, but. But you have 10 pages for one train of thought is a bit extreme. So people take less. Let's say 800, uh, let's say 400. Half a page. Half a page is one thought. So you say, all right, let's build embedding units of half page, like length. So how do we get there? It takes this long. We start chunking and that's like, okay, where do we cut? This is the big question. And it sounds so. You know, I love this because it's so basic. It's so down. It's really deep in the machine. It's a bit like oil. Yeah. Oh, this thing doesn't work because we need another oil. No one didn't. No one did care about it. But you look like, okay, um, I mean, basically we divide information into pieces and once you have the Job to divide something. You obviously ask, okay, how do I divide? Where do I divide? Here's the scissors. Where do I cut? And not um, shitting you. 90% of systems out there, what they do is okay, character window, full cut. So that means like text messages in the 90s, you maybe you remember like, hey, so the next message started with er and not and the last was ending with moth. Um, so that is bad enough, but now imagine when it comes back from the, from, from the retrieval, uh, database. It's a bit like imagine as a Google, it comes sorted by relevance. So now you have text messages from the 90s sorted by relevance. So in the wrong order. So first message and then the last ends was moth. So the poor AI has to read stuff that I wouldn't ever. But this is what is happening out there. As a standard, there are better systems. But I just realized that 90% of systems still work like this. Then you can say, okay, let's be a bit more elaborate, let's go token by token. So that basically means syllables. Yeah, so you, after syllables you can cut. Then at least you know it's mother. Okay, a bit better. But obviously the next step is let's get the word, uh, at least not cut. And then in the end a sentence or. And now we come to the pinnacle of state of the art. Let the train of thought be finished. And then cut. Okay, that's cool. That's already a lot of work going into that. But you still cut, which means you still linearly cut the text. And then imagine you have a headline and a sub headline. And then at one point, oh, you know, window is full, we cut. And then the next starts with text. And the text doesn't know that it belongs to this chapter, the subchapter, this other headline, the bullet lists don't know under which uh, text they are. That before introduces what this bullet list is about, and so on and so on. So the technical term for it is the spatial awareness of text parts towards each other. Yeah, that sounds super scientific, but it basically means when you as a human look at a page and you see it and you say, oh yeah, I directly see the structure. I see Article 1 and this and that and that and that. Uh, and it seems so structured. But what people have to know is that machines cannot see that know PDF is, is a format for printers. PDF is, PDF is humanity's nemesis, by the way, um, p. P.D. pDF is just. Is made to enslave us. Uh, no, PDF is absolute horrible. But uh, it's, it's it's so funny that everyone thinks PDF is machine readable. No, it's printer readable. And we literally printed out in our system to then take a photo and then reverse engineer what was on it. Um, but uh, yeah, so the idea is that what we feel as humans when as a human you would say, oh yeah, obviously under Chapter 3 in the subheading there, there is the, is this uh, list restoring that feeling for the AI. That is, that is what basically we do. How we do this is a different story. But that is what we, what we bring back up, that the AI is able to read and understand the text like a human and not in pieces where it just starts at fifth, this and that and it doesn't know what is the fifth.

Speaker B: One problem that I see people in software engineering tackle a lot is when let's say you're working as a front end team and then you're getting API responses. Sometimes it's just one long ass string, right? It's just one long ass and then the front end has the audacity to just write regex. It just has the audacity to write like some way of like breaking it together or as we were saying, just like doing it. That and the more intelligent front end teams, eventually they'll be like, why didn't they just send us back like a JSON? Why didn't they just send us back like a structured object? And what you were describing this whole time is really just this realization that like, instead of finding better reg ex, which is as they were saying, I think it's a race to the bottom. At the end of the day, I'm just trying to find a better reg X to parse a single string. Your thinking is more like, why don't we just make a better structure? And that's really at least my basic understanding of how Puma AI is building technology. So you're building to be able to compute all of this. The follow up question to this is more like it's meant to be accessible for everybody to use Puma, right? If I'm a nerd, which I am, I'm a coder, I know how to do a bit of code. What is my entry point to Puma AI in a sense of is it a terminal application? Do I do like a PNPM and install which I don't think that's how

Speaker A: it works, but yes, it does. It does. It actually does. I'm super proud. Uh, first of all, um, it's a perfect timing because since no, actually last night, since like from today we are live with ap. No, no, really, this timing couldn't, couldn't be better. We are live on puma.com. i'm sorry guys, I didn't want to spend €80,000 for Puma AI. Um, at one point, if you hear it, if you own that domain, come on 10K. And it's mine. But uh, I don't overspend on someone just reserving a domain. Sorry, that's, that's bad style. Until then we are pumaai.com. um, and you can just click get a free account. Thousand pages is included after that display as you go and you don't have to, don't have to talk to any salesperson. You can just sign up. This is what developers want. And before we had like pilots with you know, enterprise and this is why, you know, we only had like custom contact. But now sign up, it's free, have a thousand pages, like you know, dabble with it. Um, you have a playground online where you can live see the tree, we will get into the tree in a second. Um, and you have an, you get an API key for free. So you can just play around. There's a getting started guide, um, how you actually install it. And yes it is PIP installed Puma. I'm so proud of it. I try to, you know, grab it early and guys, we have to, don't do it. We have to get a nice pypy module. So yes, PIP install Puma and then only five or ten lines more of code that you need to integrate us in every rag or context engine or whatever you want to call it and find out yourself. Uh, what the magic is. Um, the cool thing is that now um, well, without telling too many parts of the story at once, we take in 57 file formats so we don't need pure text anymore. That was one of the learnings why we only you know, started the whole, let's say go live so late because the first, the first pilot project we did with the uh, with the corporate, they said that hey, we have super clean text. We give you the clean text and then we huge chunk it. Because my rate, my original idea was to have the world's only chunking startup like be in every pipeline, don't have any competition, everyone loves themselves and we just chunk everything. Yeah, the problem is then they said here is clean text and the first experiment was name of the counterparty print page. Cool. Exactly my humor. Um, maybe that was not the name of the counterparty but like uh, a text cleaning fuck up upstream. So okay, we need to do this differently. So we built the ingestion module. This is also why for those of you already using stuff, um, basically if you just compare end to end, what we're doing, we're doing what unstructured does. Unstructured has a focus on, you know, making documents LLM ready and chunking is more like a side dish to them. For us, it's the other way around. We have a good ingestion because we need a very good ingestion to then make the chunks. But for those of you using unstructured, it's funny, you can just try it out, like compare unstructured or us, because you can actually access it at this, like with the same architecture. Um, and we use unstructured ourselves for a long time. But uh, but now it's like, okay, both systems can documents in and uh, chunks out. Um, this is, this is what we offer for now. And uh, for people who just start with rag, I think it's, it's very easy because then you just get, you can just start with the chunks. You know, you have to. This was, for me, that was like at the beginning, the most disturbing part, when you try to get data into AI. Um, and then there's like these tutorials and you start, oh yeah, pymu PDF, just extract the text from a PDF and then you find like, yeah, but half is missing and where's the images? And what do I do? And I mean, hey, damn, I wanted my data there. Um, and then you realize, oh, but you have to cut it up as well. So you can just put the PDF in the database. If you are at this point, that's exactly perfect. Let us solve this problem. Get an account and uh, thousand pages free and write us, drop us a note. Like, we love developers. Uh, drop us a note. We can also increase that. Um, but also if you're a company, there is a way that you can just, you know, have a DPA and whatnot and pay for it. Um, no problem. We are now ready for everything. Uh, finally, and it's pip, install poba.

Speaker B: I can share one of my, like, maybe use cases, if anything. You probably share maybe some of your use cases of all the users that are using it for. I think for me, you know, the concept of diarization, some people may or may not know this word, but it's basically if you, if you do, I guess like podcasting or anything, sometimes you get a transcript and diarization is kind of just figuring out who says which sentence. Basically. Um, I'm talking to Alex at The moment if there's a transcript that gets spat out, it's like, who said which line? Kind of thing. Um, it's a very hard problem. So. So I haven't tested it yet with Pulmo AI, but maybe some of my thing is more just like the chunking. Will it help solve that problem? I don't know if it's directly correlated

Speaker A: to this Very, very funny question. Um, actually. Oh, you know, no, no, this is really. Is a very nice glide path into. Into, like, also explaining what the trick with us is, because I think it could actually work. The, the. The thing is what we do is we detect the natural structure in the text by letting. I mean, I can now tell it because we got the patent. It's, it's, it's awarded. Um, and then it's so unorthodox that if, if you, if you disconnect ChatGPT from the web, I mean, you cannot, but I mean, like, forbid it to surf and you ask exactly what, um, what we do or you. What I will describe to you now, it says, nah, this can't work. Don't do it. It's so brilliant. This is why I know we are outside of the training data. This is real innovation. Um, um, it's really like, don't do it. Don't do it. Go the classic way. Like, yeah, yeah, yeah, let's see. So what we do is we let them m. Let a model, like, we. It's an LLM family that we fine tune, but you could basically do it with everything. We let it regurgitate the same text to us verbatim, like, literally word by word. But let's say make, make, make, take. Take pauses when you speak, uh, AKA formatted like code. So we even disintegrate the whole structure. We get. We give you this line that you like in the front end, um, and then it spits out the whole text, but as a tree. What that means is it actually reads it. So it's. There's no heuristics. No reg. Ex. No. If there is a letter and then a bracket and blah. No, no, no, no, no, no, no, no, no. This is also why it speaks every language, because it literally reads the text. Like, for example, Huckleberry Finn, he has a brother, that brother has a sister, um, with red hair. And then he also has a mother. So you really see it. And this is why, actually, this could be a funny experiment. Like when you say you have the diarization, I wouldn't, I would wonder why it shouldn't detect, like, person A person B. Person A person B Person A, person B. I mean it's not exactly very deep as a tree. And it's a funny misuse of the system, but you know, I always love misuses. I always love misuses. Show me a system and I will find something funny to do with it. It's a brilliant misuse. I send me over some text and I will just, you know, pipe them through it. Um, you know, that's like, like Jens, my, my cfo. The first, I mean this bastard. The first thing. Like I say, okay guys, you know, now the internal demo is open, you can try, try out something. And I expected people to do like five page legal document, don't blow it up. Uh, should be structure. Yeah, good novel. Thank you for that. Um, so I mean, who would ever put a novel into a chunk? And then I saw, holy. It actually found something. I said, okay, that might be a bit random. And I put there's this very famous short story from Franz Kafr colony, which is basically super complex story. And then at one point the guy like, ah, uh, life doesn't make sense and he kills himself. And you see it from, from, from very far. You see it from very far in the. That the tree is going complex and then like the news, it's. Sorry, it's a bit trigger, sorry. But nobody's really like, what the. You actually see the depth of the story in the tree. So, um, honestly, I wouldn't go into. This is the ultimate narrative tool. No, but the funny thing is what I see with these examples is that it's not bullshitting around detecting semicolon or you know, lists and articles and chapters and stuff. No, it is reading the fucking thing like a human would. And this is what I think makes all the difference, that you know, we structure text by meaning and not by some little cues.

Speaker B: I know a lot of people that like, they'll, they'll have a. I don't know, they'll have a specific car model and then there's something wrong with it. And then like they'll, they'll read the manual.

Speaker A: By the way, I know this is

Speaker B: so powerful in the sense of if you're able to feed, let's say you're seeing 50 plus different formats. But we're talking about a car manual. You could Google that. Whether even like an Ikea, like building manual. What a normal user that could leverage Pulmo AI is you just find your document, feed it into Puma AI. Ah, in this case. And then it will give you a structured output. Who cares what you do with the output. The whole point is that there's a structure that lets you, you know, be more optimal in terms of trying to find an answer or even build something out of it at the end of the day. So this is what I understand why pumai is so interesting in a sense of like, why not. Why not make life more efficient as you're seeing some of the benefits. Actually, I do remember seeing the pros and cons for a lot of the. This kind of stuff, but we're seeing like, let's say reducing tokenization, for example. What does that mean? That just means that, like, it just uses less token.

Speaker A: Is it.

Speaker B: Does it mean it uses less brain power? What does that highlight over there?

Speaker A: So, first of all, um, I mean, I mean, while. While I put like many documents in puma just to look at the tree. I mean, by now, you know, at one point, uh, it's a form of art or an erotic relationship with that or I always say, you might m. I always like to say, because books are made of trouble trees. So nice circle of life, if anything.

Speaker B: Wow. And also, as a disclaimer for the people thinking about this tree, it's directly on the Website I've seen puma-AI.com if you go on there, we talk about this tree, it's just very pleasing to not only see, but to use as well. But go on, you were saying it

Speaker A: out, click, click, click. Try it out. You can. You can just upload a document and see the tree. A minute later we saw our servers don't blow over, but we send you an email when it's done. Um, and then you can see your tree. And the thing is this. But this is more like for. For the beauty of it. But you would normally. Normally don't use it. What you use is the jsons we send you for the chunks and chunk sets that you put then in the vector database. And I think the magic starts when you have like multiple documents. So. Exactly. Like your car example. So I have this old car. I have a certain generation of certain car manufacturers that I really like, and there are a lot of work. And, um. And so you find these online? Yeah, you know, like the. You know, these corners. Like the geek corner of that, actually, once you research all the valves and stuff, and I always save that and then I put it away and then you have the old, uh, someone finds the original manual and I digitize all that and then you have like, all the information ever written on this car and all the tricks, and then you Put you puma that and you put it into a vector database and okay, a hybrid full text database. And then you can just ask, okay, um, uh, it says error this and that and uh, what can it be? And now comes the thing in a normal rack. Like if you had just, you know, ingested it in a standard way, hopefully let's say you include all information. Nothing get got lost. But okay, let's assume that then it just has to find blocks where you luckily hit some triggers from your question. Whereas in puma, what would come back is okay, now it gets complicated. Imagine you have the tree. Once you have a tree, there's one thing you can do with this tree and that's a traversal path. Have now since I think most, most people will just listen and not watch, let me try to translate what I would normally do with my hands at this point. So imagine you are a table down in a document. So you are very right indented in a tree. Yeah. So you're, you know, there's a lot of space left of you because you are under the heading table for the prices, which is under a text that describes a bit the prices for which is under another heading. And so on and so on and so on. So what a traversal path is, is crawling up, uh, the tree. Like you start at the table, then it crawls. It's like, like a little cat that goes up, you know, on, on its tool, like it. But you always keep the rest with you. So you basically fold up the table, crawls up the crawls up, crawls up, crawls up. And now you have the full story from the title down to the table. All the stepping stones in between are with it. And that is one we call chunk set. That is our retrieval unit. This is what you embed. And if you think of it, this is like one consistent story. Uh, AIs love it, vectors love it, embeddings love it. Because it's not like starting something in the middle of sense and ending somewhere. But it's like in the manual for the G Wagon, uh, 500 of 1992 in chapter there, there is under oil this, and here's the information. There's a detail. So it's one vector that brilliantly points to one coordinate in the brain universe and then comes the next, and then comes the next, and then comes the next. And this is what we get out of the database. Now if you think of it, a lot of them will be redundant. So and then in the same manual, and then one below is that that would make For a lot of redundancy. But since we know where the, where the stuff came from, just before giving it to the AI, we can deduplicate. And that means in the end. Now comes the final punchline. Imagine for each question you pose to our system, we would say, okay, for each book that we have seen, we would go with the text marker, uh, mark every line and every supporting line to get there. That is important. Then take a scissors, cut them out, and then patch them together. Only the relevant parts on one cheat sheet. And this is what we give you. Done. And that obviously is between 50 and 90% less material than you would have. Uh, as with the current system, that where I always say, stop catching dolphins, you just want the tuna.

Speaker B: That is a quote that I, I cannot reuse more. Now that we have this chunk, what would you say is some of the fun use case that uh, you could do with these chunks? But basically because a lot of people will have inspiration, a lot of people will try to break it, as you mentioned already. But, but for you, what, what is, what is your ideas? What is your visions? What are, what you've seen? People use, uh, these chunks for, but

Speaker A: for many things, it's quite boring, but, but, but super straightforward. So, um, actually whenever it gets long and structured, but structured in a way like, you know, take an insurance contract, you as a human, you see it. Yeah, obviously that's very structured. Um, funny enough, you know, you're the only one seeing it, um, unless you use puma. Um, and uh, um, so just to, you know, grind materials that are complex. This is one obvious thing. And obviously chunks are made for vector databases. Yeah, I mean, you can also do other stuff with it. It's funny, we have now one case where I couldn't have got the idea. They say, okay, they have like these super high stakes, like literally one wrong answer, million, million dollars, uh, damage. So, and they parse a document and then in order for the next step to, you know, really have the right information, they put PUMA in the middle. Um, so they say, okay, can you please dissect the document perfectly for me, um, as an energy engineer, it kind of makes me sad that it's then thrown away because I always say, you know, you only have to chunk each document once. This is why at one point, you know, my dream for PUMA is to go to the original source of the document one day. Imagine every, every lawmaker, they just deliver your Puma file. Why download PDF? I mean, I want to replace PDF. Puma is like PDF for, for AI. Um, so the idea is at one point that you know, why not get the contents directly as a Puma file? So also to all the publishers out there or all the content owners, do yourselves a favor and you save a lot of energy. If not each of your customer does the same. I mean I earn more money but honestly, uh, it's just stupid if five people do the same thing in parallel. I rather sell to the guy that gives like this to the five people then um, so yeah, this is, this would be like the best thing. Um, to answer your question, if not the consumer of the data, uh, would think of okay, what can I do now with that? But if instead those who own the data say hey, you know what, my data is currently actually not really consumable for machines. By the way, funny side story, even Google would love it. No, you could, you know Google right, you know Google right now has to do full text search of websites. This is why sometimes it's so random and their, their results don't fit. Imagine Google could read Puma files. I mean I don't think that the Google guys are listening right now, but

Speaker B: just a cheeky wig. I saw that you've mentioned Puma file a couple of times. It does exist, right? Like a.com file.

Speaker A: You will laugh. You laugh when you do your tryout. You know when you upload a file, you see the tree. That's. But that's just for the visual demo. But if you get your API key, you do pip, install Puma and then you send us something, then we would say okay, it's ready and ah, what you download is a Puma file. Sure.

Speaker B: Direct replacement. Forget about the PDFs, you got your new Poma files. At the end of the day you've been solving problems, right? Like chunking. I'm not going to say you solve the problem, but moving forward there's always optimization. Is that the goal here? More optimization or more like, I guess different veins of use cases or compatibility for Puma.

Speaker A: It's cool. I actually could speed it up by honestly a factor of a hundred because before we needed to go sequentially over the whole thing. Um, and now we know, for example when we find chapters we parallelize and stuff. So we have this very funny, this very funny metric that for actually almost every File we need one between one and two minutes. Like it's super random. You can one page or 100. I mean I never expected it to be that random, but it's really like always the minute, um, whatever it is, uh, that could be faster but I can live with that. That's okay. You only chunk once in life. Like if you chunk the same file twice, you know, you just wasted energy. You should, you should respect it. So it's okay if it's slow. So you learn not to do it like uh, redundant. Um, yeah, there's two or three funny things. The one is it's still not totally solved to be able to read everything so something still stumble and especially the visual stuff. Um, that is like, you know, when someone has a bad handwriting and so on. Then in terms of cultures, um, I love. I mean I honestly think it can already do it. We are good with, you know, top to bottom, like Chinese and Japanese and stuff. We're currently running an experiment with Arabic because it's super funny. Arabic is a right to left language. But we are indenting from left to right now. So we basically have the schizophrenic tree. And maybe it works because secretly the AI also like the systems in terms of code. They think from left to right. I mean, talk to Arabic coders, it's like commentary is from right to left. But the code is like what the. But I want to. But I need to find out if it blows up or not. The experiment is running today by the way. It's super funny. So like, you know, I really, I don't. I just don't want to tick boxes in terms of language and culture. So. But like. No, no, we can do this actually also for stuff that uh, you know, is something totally different. Um, that is one thing. Then I have this thing. Um, a secret dream of mine is uh, to make Excels explainable. So uh, when we import an Excel, we don't import only the values like okay, the table, but also the formula and then the formula traces. So we can basically say, hey, um, you know. You know, like in investment banking they have these, these big spreadsheets and they think, yeah, obviously it's um. It's. How do you call it? Uh, um. Like it's not traceability but account now. It's not accountability, but you know what I mean? Like you can basically do the whole calculation yourself and then explain each number. Yeah. Have you ever tried it? Like, uh, it's so complex and no one knows what they're doing anymore. But now you could finally ask, hey, my banker sent me the spreadsheet. Can you explain to me how. How he got to the P L last year? That's kind of. Because that's actually useful. I mean it's not sexy, but it's useful. Super useful. Because honestly, even I, I, I, I get, I get my, my, I get my balance sheet and I say m, yeah, it looks very good and I sign it. We honestly, I cannot calculate each number myself, but with this I could. So it would actually help people also not get to get, get, get screwed over with this stuff. Uh, yeah, that's just one thing. And then the last funny thing is, uh, but this is, I mean, this is more the fun stuff and not really useful is importing videos. Um, because I, uh, always have this dream of making AI watch YouTube for me. So I, because I, so I can save the time. It's different with podcasts. With podcasts, it's actually funny podcast. Listening to podcasts, I'm very auditive guy. I also have like 500 audiobooks and stuff. Yeah, for half blind it makes sense. But like watching YouTube videos where someone is like, yeah, you installed this by blah. And you're like, okay, come on, I can do this with a transcript and keyframes. And I'm looking forward to my AI being able to watch YouTube for me. So that's the fun project.

Speaker B: Maybe this is coming back to Alex. Alex yourself is that you've gone through so much and you're going to go through so much more. At the end of the day, what was some, like, Guiding Lights or like Northern Stars that you've always like, kind of looked up to whenever you made decisions, whatever, Whether it's like the entrepreneur side or even just the technical side of learning things. I think a lot of, lot of people think about them, think about this, about themselves. Right. Very introspective question, but I came on saying that, like, just listening to a story, it's been so inspiring and so interesting. What are your Nerf stars? What are your guiding lights?

Speaker A: Well, I read a, uh, lot of the, well, the classic biographies we all know. And I have to say, I mean, I. Why many of them took some random turns in their later lives. Um, it's still inspiring. Like the origins of being 00 to 1 for Peter Thiel or like Jobs Musk. I mean, you know, the classics. It's not nothing, nothing super new here. And I wouldn't say I, I ever had any of the what would Elon do or what would Steve do? Kind of moments. Um, no, no, but rather when you just digest all these stories, you kind of. I don't. I abstract. I triangulate a bit myself. But what I have as guiding posts is like Golden Rules. Like you could say, um, what is the name again? Like, first principles. Um, and they Are a bit random. They're not like what you expect now. But the first is like, nothing is more expensive than ego. And this rule is so good at sometimes. You know, once at fair, I said, okay, we will now burn 10% of our marketing budget for one super bowl spot. We actually. I mean, only the German, uh, the German transmission of the U.S. super Bowl. But in the last break, we had one spot. And, uh, and there was like, crazy expensive. And I knew I just wanted points. It doesn't make any sense, but I want it. But I. I face the fact that this will be expensive. This is ego. But in some. At some points also in negotiations, you have to know that, hey, I can go to. I can easily give in if I don't. You know, it's just ego hindering me. And knowing that ego is the most expensive thing you can have, um, helps a lot with avoiding bad decisions or taking some decisions. But like, you know, uh, within. With an upright phase. That's one rule. The other rule is every revolution eats their kids. Like, uh, you know, start. Yeah, yeah. Love open source whatnot. At one point, you know, revolution will bite and. And, uh, the. The. The Guerreyas will become autocrats. Um, so that. That always helps to keep in mind that, uh, you know, our revolution will end at some point. Um, and. And then there is, um. I don't know what was it again? Um, but, um. Well, it doesn't come to mind. But, you know, see, there's not like a list of one to five. There was like two of the. But like, let's say golden rules in life that always, you know, whenever I experience. Yeah. There's also like, sound. Sounds banal, but, you know, after rain comes sunshine, but when the sunshine comes, don't throw away the umbrella. Um, this is about one of the more banal ones, but no really, like, stuff like that. If you have like a. A mental toolbox of like 10 of them coming up that kind of, you know, helps you triangulate. It's, um. And. And that's what's. What's guiding me. Um, but also a bit like, I'm also a bit crazy. I mean, if in trouble, double. I'm honestly like, bring it on, bring it on. When I really think, like, okay, man, we're close to dying then. Yeah, but we are not dead yet. Bring it on, baby.

Speaker B: You can end it better than that. At the end of the day, we got. We got all the idioms. We got all the fighting spirit until the end. And even though I keep on saying this is the Beginning of the journey. There's so much more to come on that. So for me, after hearing all those kind of stories, just gets more excited. Definitely people go check it out. But the real thing is where could people find Puma AI? What's the outlets, websites, online socials, where can people find more?

Speaker A: So, yeah, like our, our main outlet is puma AI.com. um, and obviously we are active on LinkedIn and X. Uh, we will now be a lot more active than before because we were, I mean some people, some people think we're like a defense stealth lab or whatever, but because we did like only these 10 pilots and our only button was the book, a demo or contact sales. Finally. It's not anymore. I hated companies that you can only get to through sales. No, no, you can just click and get your account, um, and get in touch like. Well, obviously you can also write us a good old email. Why not? Um, but yeah, I hope you find us on LinkedIn and uh, X as well. Um, and uh, we will start being more active on Reddit. I didn't do it until now, but now that we know, um, developers can just sign up and try it out. Obviously I will reach out with the, you know, open openvisor, like, hey, I'm the founder of Puma. We built something cool. Um, here, get to get a free key and try it out. I want your feedback. So we didn't do this yet, um, but this will start now because I really want the developer outreach. Uh, until then, you know, I hope you hear from us, uh, yourself. We're doing a lot of SEO on all the keywords that when, when you search for sentence splitter or recursive character text splitter, like the LangChain library. I really hope you directly find the page I wrote for the last two weeks. Um, so, uh, if you have the problem of chunking, you will find us. Otherwise, puma AI.com.

Speaker B: all the links will be shared in the description down below. Please definitely go check them out. But on behalf of myself, Alex, it was an incredible story listening to all of it. Thank you for being on the show.

Speaker A: Thank you. It was absolutely cool, was a lot of fun.

Speaker B: Uh, we love hearing that. And of course maybe you'll be back on the show one day with the next expansion of POMO AI, if anything. So until then, thank you everybody for listening. I will also catch you on the next episode of podcast written by a software engineer. Bye. And that was Dr. Alexander Keem, founder and CEO of uh, Puma AI. All the links are shared in the description notes below, go check them out. If you would like to be on podcast written by a software engineer, reach out to us. Just email contactairetsu.com can't wait to have you on the show. Once again, thank you so much for listening to Podcast Moved by Software Engineer. Don't forget to hit the Follow button on your podcast app and leave us a review on Spotify and Apple Podcasts. It's completely free, and if you want to tell your friends about Podcasts Moving by Software Engineer, that's totally cool as well. Anyways, I'm parentsu and you've been listening to podcasts proven by Software Engineer.