FikaAI Interview with Michele Catasta
Recording: FikaAI Interview with Michele Catasta
Hello everybody. My name is Alexi Kraov. I'm the AI community architect at Neo forj and founder organizer of AI by the BB RI meetups. With us, we have Mika Kadasa who is the president of Reb. Correct. And Mikuel just was on a panel. What was your panel about? I think the panel was about how to build AI applications today, especially agents. That that was the focus
And so Reblelet agent is number one I think application builder thing and like we've seen you evolve from basic things to an iPhone. So tell us what's like the most common way for people to build apps with rapidly these days. We've seen a lot of variety on what people are building with our hen. Definitely a lot of full stack applications including database integration hot and so forth. We're also seeing a lot of data dashboards. People find them very useful to process data and put it in front of customers. And I'm very excited about internal tools in companies because you know they allow you not to buy very expensive SAS create a product that is very easy to use for everyone and we're getting a lot of traction there as well today. Yes
And like if you are on X you see Amjad always posting like this guy just built an app over the weekend which earned a lot of money and like this guy was waiting for some team to build something he got tired and build it in replet. So like do you see a lot of the stories coming all the time and I think we repost them because we're actually very excited about them. You know it's not purely marketing is the idea of feeling wow this product is really enabling people to create software 24/7. Yes. You know like even when they have you know some downtimes during the weekend rather than you know doom scrolling on social they spend time building a replet. That is such a you know wholesome mission that we're accomplishing. Absolutely. And so like what what really excites me you're really democratizing it right like I think I'm said he's one he wants to make 1 billion developers right and I just you know stopped by the talk from your de guy yeah and Matt is like showing a slide which says like what is the bugging right like what is API call what is you know 404 right so do you really see people who did not have coding experience getting into this and basically using AI to learn how to do this absolutely Absolutely 100%
We have seen people with no coding background whatsoever becoming successful at creating applications and not only they are successful at creating but in the process they also learn like our agent is very chatty. It explains every single step that is taking and we do that you know on purpose just to make train your you know skills in creating code with the agent as time goes by. So the more you build the more proficient you're going to become using it. Yes. Do you keep track of what most common questions people have? What most pitfalls they have? Like basically you can probably make a course based on the most frequently asked questions. First of all, we are about to make a course with deep learning AI exactly covering all the pitfalls and the successful, you know, prompting strategies that you can follow and yes, we do a lot of datadriven analysis and in a sense we prioritize the road map of what we build on agent based on the both successful prompts and you know whenever users get stuck in the process and you know it's a journey definitely the agent we have today is the worst agent we will ever release. You know, every day we release a better version and we're going to keep doing that for long. Tell me a little bit about this evolution
I remember you partner with lang chain. So you basically basing on langraph. Correct. Right. So what does langraph give you? So langraph is a great starting point to think about agents in a in a principal way rather than trying to reinvent the wheel yourself and it comes with also a very good side benefit of this tight integration with Langsmith. Blackmith is an observability framework for agents and I was you know very adamant about having observability from day one in the project because agents can fail in such a so many wild different ways that if you can't really track what is going on you're really never making progress you're always going to be adding regressions so that that was like a very useful tool that helped us since before we released a v1 back in September. So this is a thick so the series about people in there. So I just want to ask you a little bit about your personal history, right? And I was digging through my emails as you know, you know, I mailed you and then I found that we mailed back in 2015 when I was in Florence at the conference and you were a graduate student and you were doing Scola and you asked me about some data from mechanical tur because I used to be at Amazon before
So tell us a little bit about your own evolution in software engineering, how you came to be president of Rebel and what you did before. I started as an open-source software developer. I've done a lot of work under the Apache umbrella. I had contributions to Adoop, Lucin, Solair. That's where I know I I I trained my open source muscle over time and then I became very fascinated with what was called machine learning back in the days and the reason is had the tools to manipulate data and I kept asking myself what can I do with it aside from processing it. So I you know made my entire career bet on machine learning and I started you know to work on crowdsourcing a bit because as you know back in the days that's how we collective labels for supervised learning right and then over time you know deep learning accelerated very much so I became more and more interested the moment that transformers became you know available in open source I immediately embraced them to work on basically language models for software creation you know the rest is history I've been doing research on that at Stanford I was at Google working on Palman and Ponto the LLM generations right before Gemini and then the moment I felt models were becoming powerful enough to create something like rapid agent then I decided to focus more on agent building rather than purely training models this is the last challenge I've been going through in my professional career and it's a lot of fun to be honest I think it's amazing right like at every juncture you made the right choice you went where the gradient right is the most efficient you follow the gradient you follow the gradient correct the gradient so but what's interesting to me. So, I've been a Replet user since way before it became an company because it was a niche little terminal, right? It's a ripple. It's known for functional programming geeks
And so, it used to be basically you spin up an environment for your okamel or host or something because it's hard to do, right? And so, how did you connect with Amjad like how did you move basically replet from the little terminal which where you run ripples to an AI agent like what's what was the evolution? How did you converge with thinking of atlet before that? I think the mission that Amjad had in mind back in the days was always the same. Empowering the next billion software creators. He started the company with the assumption that making the development experience easier would have lowered the barrier of entry which is true. But what we found out over time is that we needed to drop that barrier of entry by orders of magnitude rather than gradually. So when we initially met we started to brainstorm even more about agents even though they were not even called that way back in the days and the moment we felt we could have built rapid agent we pushed basically the entire company beyond it and then we made it happen like literally 6 months ago. So again we're following the gradient you know the curve was a bit less steep back in the days and the moment we could make it happen we really created a product that billions of users could be you know using already today. Fantastic. And so the last question is like the developers kind of spectrum
So there are newbies who know nothing and people like you who know functional programming and right like everybody in between. So what like is replet only for newbies or are there sweet spots for experienced developers who want to improve replet agent itself? There are definitely sweet spots and we have a lot of expert developers on rapid today. The reason is even though you will be capable of building a full stuff web app on your own time is correlated with money. So if you if you feel that the agent can make the app you know with one/10enth of the time it will take you to do it on your own of course you will trade that time you know for money and that's what a lot of developers are doing are creating the initial prototype on rapid maybe after that they're taking it out they are either deploying it or they're further developing it with their own skills but I think we're empowering everyone to be more productive some people will hit the ceiling because their skills are not there yet and some others will take whatever the agent created and evolving it. Nice. Actually, one more question after this. We are in Neo forj where graph database lang graph is a state machine for agents. What like what the best thing you like about graphs and how are graphs used in in the replet? I think graph is an amazing abstraction actually also done rearch research on graph neural networks
So it's something I've been very passionate about for a decade at this point. The reason why it's fundamental for an agent is because you're not building a you know flat uh or linearized agentic workflow. There are a lot of different decision points that you could be taking. And the reason why you represent what an agent does as a graph is because you have a feedback loop. You have an environment where you take actions such as you edit your code, you execute it and then you you you debug it. So basically you go back to step one and then you head the code again, run it and debug it. So graph is the only type of data structure that allows us to really capture all the different step that the agent can be taking. slightly more advanced than what has been built in the past but definitely is the good starting point and that's why I recommend people to at least take a look at lang graph documentation is the is the right starting point the right obstruction and this is all open source lang chain is an open source project so open source so you guys with new forj contributed a lot in open source so all our develop open source so we are happy to work with folks like you to build apps on top of we love open source and rabbit as you can imagine fantastic thank you Miguel appreciate Of course
It's been a pleasure.