Code isn’t the only thing causing your production failures

**Ryan Donovan** (0:08)
Hello, everyone, and welcome to The Stack Overflow Podcast, a place to talk all things software and technology. My name is Ryan Donovan, and today we are talking about code and production, specifically the increased issues with AI code and production, and how some of the AI approaches are not gonna help you fix those problems. So my guest for that is Anish Agarwal, who is the CEO and co-founder of Traversal.
So welcome to the show, Anish.

**Anish Agarwal** (0:38)
Thank you for having me, Ryan.

**Ryan Donovan** (0:39)
Of course.
So before we get into the topic today, we like to get to know our guest. Tell us a little bit about how you got started in software and technology.

**Anish Agarwal** (0:48)
Absolutely. So I've been in the US about 15 years. A lot of my time has been in technical institutions. So I was at Caltech, I went to grad, studied computer science then, and in my senior year fell in love with machine learning.
It was kind of coming up as a topic then, and when I saw it, it felt like this beautiful intersection of computer science and statistics that just made a lot of sense to me and felt it was going to get important. I spent a few years in industry and saw how much machine learning was starting to already have an effect on enterprises, and so I felt I wanted to understand it better than I did at that point, and I felt a grad degree was the way to do so.
So I was fortunate enough to get into MIT's Ph.D. program in computer science, and spent five wonderful years there studying AI and ML, and quite broadly, but then I focused on a few different areas, one of them being causal machine learning. So many people have probably had the phrase correlation isn't causation. These AI systems are very good at picking up pretty minute correlations in data, they're not very good at picking up cause and effect relationships. So my research is how do you get these systems to learn cause and effect relationships from data programmatically. And the second area of research I had is an area called reinforcement learning, which is fundamentally about how do you search large spaces effectively, in an efficient way. And so those are two things to study and they're actually quite deeply connected, you can get into in a different time or at some point in the talk. And so yeah, I mean, I loved it. My goal was to keep doing that and continue to research. And I felt being a professor is the best way to go express that. And I was lucky enough to get into Columbia's faculty, which is what brought me to New York City. But as I was making the transition was when, you know, all of these incredible things with chat GBT started happening. I mean, I'd always played around with all the stuff that OpenAI put out, but this felt kind of different and just like we had hit a phase shift and what's possible. That's what made me decide that I think right now some of the most interesting things in AI are happening in companies. That's where a lot of the most interesting research is actually happening. And I wanted to be part of that.
So that's kind of what got me back into company building. That's kind of my broad experience with software. I can talk about how we came to the specific problem we're dealing with the traversal and why it's really interesting. Maybe that gives you a little bit of my background into software and technology.

**Ryan Donovan** (3:03)
Like you mentioned with that phase shift, it's one of the outcomes of that phase shift. It's been a lot easier to produce code. You still have the issues with running it. And I think a lot of the code generated by AI is not as high quality as other code. What are you seeing with code and production issues?

**Anish Agarwal** (3:24)
Yeah, it's really interesting. So when we started the company about two and a half years ago, we started in some ways without an idea. We were looking for something at the intersection of our research collectively, me and my co-founders in CausalML and RL, and how that fit with agentic systems and LLMs. In that time period, we were also playing with GitHub Copilot. That was the thing two and a half years ago. And we started playing with it and it seemed really interesting. It was good at doing these autocompletes for you, like localized autocompletes.
But we just put ourselves in the mindset of what happens if as this grows and this gets better and better. We were fully bought in that this is where it's one of the major use cases of AI is going to be here.