Part I: From MIT Researcher to YC Entrepreneur. Building Workflow Stack with AI w/ Bernard Aceituno

Speaker A: Hi everyone. Welcome to Masters of Automation. Bernard, it's a pleasure to have you today with me.

Speaker B: Pleasure to be here.

Speaker A: It's great to have you here. Today we're at mit. Uh, it's actually where you and Tony met. Yeah, indeed, for a long time ago. But just to kick things off, you have an incredible career. I mean you started as a researcher, right?

Speaker B: Yeah.

Speaker A: And then you had a. Before or before you're start focusing on Stack AI, you're another startup as well.

Speaker B: Yeah, but many years before.

Speaker A: Many years.

Speaker B: A different life.

Speaker A: Different life. But tell me more about like uh, how, where did your career start and then where did your passion lead you? To Boston.

Speaker B: Yeah, so I have a very eclectic career. You know, I'm originally from Venezuela. I was born and raised there. I grew up in a home where, you know, I was always invited to pursue my passions and my uh, interest. And I was very interested in math, physics as a young kid, living music and uh, when the time came to go to college, I couldn't leave Venezuela at the time, so I had to go to the university in Venezuela for a long time. I had to start my career there. I always knew I wanted to be some sort of scientist through that time. And so I studied um, computer science and electrical engineering at the time, which I knew would give me the breadth of opportunities to kind of have a background in math, physics, but also in computer science. And um, at the end of that journey, I, through all the political crisis in Venezuela when, uh, most of the universities were Catholic semesters by, you know, through protests and stuff, I decided, okay, well I have a lot of free time. I'm spending six months at home, I might as well start doing research. So I started doing. I kind of like learned everything on algorithms, applying algorithms, machine learning, starting since 2012. And since then I've been involved in the field. I started doing research kind of like on my own at the time and published my first papers with the help of my undergrad, uh, advisor at the time who was a slightly old professor. And slowly through that I built a research career that led me to come to MIT to do my master's degree where I, you know, started doing planning algorithms for robotics, uh, which is slowly turned into more like applied math research towards geometric problems that were very relevant to machine learning optimization. After that I did my PhD. I know I spent a small stint of time in Facebook AI research when I worked on the first type of transformers, mostly applied to reinforcement learning. You know, it wows me to see how much things have evolved since then, which was 2020. It was just four years ago until now. Uh, it's not like I was in a small lab. I was in probably one of the most well funded research labs in the country. Just with the pace of four years, it's been, uh, amazing to see how much it has evolved. Fast forward, uh, Tony and I became bigger friends at MIT since we started. Yeah, we both knew that we wanted to at some point leave the academic world and go and bring this technology to the outside world. We both have been fascinated by automation for a long time. We both had been founders before. He did some work on, you know, devices for safety and like writing. I did some work on FinTech back in Venezuela before I came to MIT.

Speaker A: What was that about?

Speaker B: That was a very, uh, long time ago. Essentially, like, well, you know, I don't know how familiar you are with the economies of America, but back then Venezuela had a specific problem which was that it was illegal to do transactions in foreign currency. And the country was going through an inflation hyperinflationary period with high devaluation, meaning the currency lost value extremely quickly. And we had a ton of inflation at the same time. So it was impossible to live with the current for the franc with local currency. And it was illegal to do transactions in foreign currency. That means you can use US dollars or euros or any more stable resource. Back in 2017, this law didn't apply to cryptocurrencies. So me and my friends who had been kind, uh, of interested in fuel for a while, we said, well, it could make sense to help companies, small businesses, do transactions in cryptocurrencies, at least with stable coins or with coins that quality to make fast transactions. So, you know, we started working with that. We had a small app and you know, we grew pretty quickly actually. Ah.

Speaker A: How many users did you have?

Speaker B: We had 5,000 businesses in Venezuela using us at one time and more than, I know, maybe like 100,000 people kind of like using at uh, a time to the transaction. We had some regulatory issues around the time I got into Venice. I got into mit, among other schools to come to grad school. And I kind of had to choose, okay, do I want to keep pursuing this, which has like a very low chance of success where I'm going to get jobs with government. Well, I know I'm going to have to deal with this issue, or do I can go and pursue my academic career and then try again in Silicon Valley or something. So I took mit. Uh, I always knew that I really enjoyed the idea of building something, shipping it and getting people to use it. And while I was spending my time here, I kind of followed that mindset through the years.

Speaker A: That's incredible. I mean when you think about it, you solve not only an opportunity at the time, but you also saw a big problem, uh, both in the country and the way people live in place. Um, then making the decision to pursue your career is definitely a tough choice still, right?

Speaker B: Like absolutely.

Speaker A: The entrepreneurial dream behind which also like ended up being. You ended up becoming entrepreneur anyway.

Speaker B: Yeah, to be honest, like I always knew I wanted to be a scientist. Also like I was on the fence. I still have a deep passion for science in my free time.

Speaker A: Tell me more about that.

Speaker B: I know like I so I love mathematics, I love physics, I love algorithms. Uh, I read research papers on a daily basis still. Even though I might not apply in for stack AI, maybe completely unrelated stack AI, I'm still always kind of like trying to stay on the fence on the Lewis guy. People still email me about my research, they ask questions. So I, you know, I'm always on my email kind of like looking to help them kind of like understand the algorithms I made. And yeah, like I m miss the days where, you know, that was my full time job. You know, like now I'm doing something where I can see the impact in a much more immediate way.

Speaker A: Yeah, I mean it is interesting because it's like in a way. So you did the start, you did the startup, um, and then you came to MIT to pursue your research and what was your research about?

Speaker B: So I did all sorts of things. So uh, you need to go into a lot of detail. Uh, I would say that I work in two specific fields. I work in planning and optimization and reinforcement learning maybe like so again as a subset of planning and I work heavily on algorithms for planning and how this could be uh, designed to have certain properties and certain geometries. You know, if I feel into too much detail, we'll get into a two hour conversation that will get, that will make your audience leave the call. Yesterday I did a lot of algorithms and uh, optimization and then that had some applications to reinforcement learning. And when I went to, when I went to Facebook, AI research, that actually was kind of where I started pretty much my efforts taking some of the algorithms I made and like adapting them for the AI world.

Speaker A: So let's talk about then because what is very interesting is especially having a research background and love for physics, mathematics. Then over time in the startup life translate into more meetings and then customer support tasks and like more customer success type of things like that becomes like an everyday. Especially like you're transitioning from a researcher to an entrepreneur. What were the some things that shocked you? Right, like that was different.

Speaker B: That shocked me. Wow, that's a good question. I mean I knew that I was a vegan entrepreneur before and I knew that it was just uh, you know how I was eating mud all day. That's how they call it, how they say it. Uh, so I knew that it was going to be a lot of hard work. I know it's a lot of doing things that don't scale at the beginning. And I always felt comfortable with the idea. Even at my research I was always into like doing things that don't scale approach to life. That was something that I realized. Something that shocked me at the entrepreneur is how important is to be in an environment where you are surrounded by other entrepreneurs or people who live like entrepreneurs. And I'll tell you why. And that's actually one of the reasons why I think it's so important to be in a set for many people. When you live a life where people around you are not working 24 7, are not making little money, are not embedding their life into a long term project, you get demoralized very quickly. Uh, because you don't have that vision of everybody around you. They have a job, they take the weekend stuff, uh, they can go out for dinner every day whenever they want. They have vacations and you see, okay, why am I doing this? Am I crazy? Am I wasting my time? And that demoralizes you a lot. And the tourist at a startup is very, it's one of the hardest things you can do in life because both is very, it's very intense. It's as a very delayed reward towards whatever you do. And you need to be very driven by the passion to solve the problem you're solving or to satisfy the audience that you're satisfying. It's very important to be in an environment that allows you to be in a mindset where you can dedicate yourself fully to that problem and not feel the urge to give up or to question your motivations for the problem. And with the wheels and customers is that everybody is living this entrepreneurial life and even if they're not, they understand it. So even when you hang out with people that may be doing sales in, I don't know, Oracle, they understand, oh yeah, that's the start of life. Everybody's like this. Whereas, you know, sometimes I get my friends who may be working, I don't know, in a, uh, you know, some sort of financial consulting, which also work. But for them, they say, okay, but I also work a lot. What do you mean? Like, uh, you kind of come out tonight. It's like, well, different

Speaker A: in a way that it allows you to actually feel motivated even though, like, you stay in and heads down. Absolutely. And working. And to that point. So walk me through, like, how you and Tony, like, uh, your researchers, and then you said you stumbled upon Stack AI. Like, how did that journey come.

Speaker B: Yeah, uh, that's a very good, Very good question. It's a very fun story. So Tony and I essentially like became. We became friends at the beginning of mit. We actually became friends in AI conference in Montreal where we both met. It was funny. I always tell people go to the poster sessions in the AI conferences because that's where I met Tony in the poster session. I saw his post. I told him, hey, I heard your name. And we just kept talking, became friends. And you know, we saw each other occasionally here in Cambridge. Then after Covid hit, everybody was, uh, locked down. Nobody wanted to go home. Most a lot of people were working remote. A lot of our friends graduated and left. Uh, people were very interested in hanging out and like, uh, you know, just talking and meeting other people, except for Tony and I. So you became very good friends over Covid. We became super close. We also realized we were neighbors. We live like a block away from each other, which here in Cambridge is pretty small. And, uh, we both knew and realized very quickly, oh, we have both this passion for algorithms, mathematics. You know, he was very interested in robotics and computer vision, and I was very interested in planning algorithms and optimization. And like, through that path, you kind of became very good friends. And we both knew, we both had the same background and we both talk about, should I go? You know, eventually, after the PhD, we probably both want to be in a startup, uh, and one of different paths. And after a while we just said, hey, you know, you want to be in a startup, I want to be in a startup. Let's begin a startup. Let's be a startup together. You know, we kind of like doing it. Probably we started doing, taking it seriously, this idea in May 2022 when we were about to graduate the PhD when you. It was becoming more realist. Okay, we have six months before we graduate our PhDs. Or we could graduate in six months. Let's actually take it seriously. What do I do? Okay, because, you know, we are done playing around with ideas and whatnot. We say, okay, for sure. We know that we, you know, AI is A way that is coming. And uh, we both have invested significantly in this field. We know what are the problems that kind of like people building tools with AI face, especially when it comes to bringing AI to the industry. That was always like our motivation has this entire breadth of power to be a system of analysis to bring insights to automate tedious and unnecessary work. And the reality is that most AI at the time was just living in a jupyter notebook. Only a data scientist that could go collect a data set, train a model and then deploy that model in some complicated cloud could really like apply it. And because of that data, AI was always limited to that data scientist, research world analytics, world of the enterprise. And it was never going beyond that scope. Very rarely, sometimes very, very specific, application specific setups. Uh, we said, okay, maybe we should like focus on solving that problem, which probably is a big thing. And we said okay, how can we towards that our first question was, okay, why are people struggling to bring more AI models? And we thought about it, especially the data scientists and we weren't very excited about foundation models at the time, but we knew that it was potentially fine tuning them, training them and having the ability to build your own models. We actually had a lot of experience before building uh, foundational computer vision models for an idea we were trying in the past. So we said, okay, this is for sure a problem. And the biggest one for us was data. Okay, how do you get good data? And at that time we founded Stack AI as a tool for taking a data set of images or text or whatever and cleaning it up and giving you the correct, the best labels to train your model or fine tune your model.

Speaker A: Interesting. So it was more like um, preparation tool.

Speaker B: It was a preparation version control. It was a very mature tool. It was focused towards training models. We took the idea, we built a prototype, I reached out, we reached out to a bunch of people. We got a few clients and from those few clients, you know, we actually got people trying and they liked it. So we said, okay, we have our together it's a pattern Y combinator. So we applied to it. It's actually the second time we applied before it was kind of like us playing, but we applied seriously say okay, we actually want to do this idea. We have customers, we are very convinced that, you know, like uh, this idea of anything is going to grow. And then we apply. We got in and you said, okay, we got in. Maybe we should grab with the PhD.

Speaker A: How was that process like, so you applied to IC and then like, like what were the some of the things that they asked and, uh, the initial

Speaker B: questions they asked in interview are very straightforward. It's like, why are you building? Why do you know this is a problem? Who are your customers? Which customers have place? Anybody trying this? Um, how much time before the project is ready, you know, to like, actually, like grow and scale? Very basic questions which, I mean, you can follow very quickly in the first of them, where are you building? And then second one, who are your customers? How you know they won this? Like, it's very easy to follow through. The waste interview is 10 minutes. So it's very easy. But we got through it. Um, and we said, okay, we got into YC. Shoot, we're still in the PhD and YC is in two months.

Speaker A: How long did you have? Like, until we had two months.

Speaker B: But, you know, it wasn't guaranteed me. Uh, I was in a position where I had most of my research set up. I got very, I would say lucky. But, you know, thankfully I had a few breakthroughs earlier. My PhD allowed me to kind of, uh, make my final year much more about aggregating the work that I had done into kind of like some, um, key results rather than kind of like still trying to pursue those breakthroughs. Yeah. And that helped me kind of like accelerate my graduation more. Tony had some work to do, so he had to work a lot to finish. I can't do this. Okay. Bernard will code the first version of the platform while Tony finished his Ph.D. and you know, works on again the branding landing page. Then we'll switch and we'll switch and, you know, we both printed to finish the Ph.D. finished December, two days apart. December 15th, something like December 17th. December 17th. 19, actually for him. 19. And then we said, wow, we're done. We took 10 days off because, you know, it was New Year's and I got, you know, like, we also had a higher lease here still. So 10 days off. And then we moved to California. We took a flight, arrived to California and started working stack AI right away.

Speaker A: And how was that? Like, so you came right there and then you had a product that's more on, um, to prepare a foundation model with better, uh, data labeling.

Speaker B: Yes.

Speaker A: So what made it more through the YC program for you guys to realize that, oh, wait, let's maybe pivot it a little bit more, like, take a different direction.

Speaker B: So at that time we were serving more like data teams and startups. And, uh, we took a big bet at the beginning yc, we said like, okay, like, there's an extreme demand for fine tuning language models. And building custom language models and ChatGPT had just come out and you know like people, people were excited about it and had to make it. They said okay, let's build the product where it's more to focus on language models and NLP and uh, to go towards this task. And we started getting a lot of amount of customers that wanted this or they thought they wanted this. So we really, we took the product, we shifted towards ninwas models. We made an announcement online, you know M hacker news, you know, internalizing it or not and we got some emails. A few companies wanted to try it. We reached out to people, hey, you know we have this tool and we got like our first four or five customers on like that we just might this NLP idea. And I said okay, this is worth pursuing, let's go deep. Okay, so what do they want? They started working with them at the beginning because the product is kind of being built while they're paying for it. You're always like very on top of them and you're kind of guiding them and you're almost like when doing consulting that's normal because you're like giving them a product that kind of half works and you're like optimizing the product that they want. But through that process and those first two months Y Combinator a uh, bigger decision we had was that most is what these companies want to do was to take some foundation model and enrich it with the tools and data that they have in their business in order to automate or help access information in some process that they had. And uh, really towards that you don't really, to achieve that you don't really need to fine tune a model. In fact fine tuning a model will like ruin a model towards this task. Like you can make m. You can steer it towards uh, following a particular language or a particular format in the output but you cannot really solve this task by fine tuning a language model. What really was important was to compose logical kind of like a structural frameworks and workflows where you will retrieve data, hand it to a language model Though it kind of became rag at the time. Back then people didn't call it rag because there was only one paper about the topic and it wasn't very popular and it was talking anyway. It was very funny and I guess how the term came up and almost the.

Speaker A: Because I think in the beginning especially they're like the different type of users. Like there's the, there's the business users who are more on to like um, like low code style app and then they're more technical users who will maybe feel more comfortable encoding.

Speaker B: Yeah, yeah, but when, when you, when

Speaker A: you were getting those first customers in, how was the Persona like?

Speaker B: Yeah, that's actually, and that's actually something we tried a lot. Once we realized that what made sense was having more of a workflow builder tool that could like achieve this task and release its version, we got a lot of interest. Like the moment we released our first launch online, we got 10 customers in a week. And, uh, we said, okay, this is definitely worth pursuing. The set of customers that we initially got were more on the executive business side of things. And that was very interesting because we said, okay, there's definitely a potential here for the enterprise to build internal tools and a potential there for the startups to kind of like build prototypes and a potential for the small businesses automate small processes. I weren't very sure which ones to pursue. The first batch of customers were kind of like in the startup and enterprise side of things. And we saw the differences there. There were developers that were prototyping something. There were companies that were building some internal tool. Slowly we found that the business people that built an internal tool and use it actively, once they have this tool working and they gain value from it, they never leave it. Developers, when they build a prototype and they release it, they move to code it themselves. So, you know, it doesn't really make much sense to build a no code tool for people that want to code. With the, uh, through that experience, you said, okay, our focus should be as that business user. And then the event was, okay, do we want to go to enterprise? Do you want to go to small business?

Speaker A: Friends, a technical hiccup interrupted the rest of our recording. Rather than serving you an unfinished conversation, we're pressing pause here. Bernard and I will jump back on the mics to finish the story. Thank you for listening to us.