Greptile: AI That Actually Understands Codebases

1. 0:00 – 0:18

   Intro & What Greptile Does

   1. BFBrad Flora0:00

      Hi, I'm here with Daksh from Greptile from the YC Winter '24 batch. Thanks so much for joining us.

   2. DGDaksh Gupta0:05

      Thanks so much for having me.

   3. BFBrad Flora0:06

      All right. Um, so what is Greptile?

   4. DGDaksh Gupta0:09

      So at Greptile, we're building AI that understands large codebases, and software teams use it for things like reviewing pull requests and answering questions about their codebase.

   5. BFBrad Flora0:17

      When you think

2. 0:18 – 1:00

   Why Code Review Matters

   1. BFBrad Flora0:18

      about the value that a company or a user gets from Greptile, like how-- what is it? Where are they finding utility?

   2. DGDaksh Gupta0:25

      It's a couple of things. For software teams, especially senior members of the software teams, tend to spend a lot of time reviewing pull requests and reviewing code in general. And, uh, humans are imperfect at this, and it is also a very time-consuming process when that time is maybe better spent doing more creative things. And so we take away a lot of that time, reduce the amount of time it takes for code to get reviewed, and we usually reveal many more bugs than a purely human-driven process would.

   3. BFBrad Flora0:50

      The customers that you work with, um, do they want to have nobody reviewing code, and they wanna ha- entrust Greptile with that a hundred percent? Do they wanna have some

3. 1:00 – 1:40

   Humans + AI in the Loop

   1. BFBrad Flora1:00

      people doing this as well? How do they think about it?

   2. DGDaksh Gupta1:03

      So I think virtually all of our customers and us, we both agree that, uh, a human should be involved in the process. I think code reviews, as a part of the software development life cycle, serve a purpose greater than just detecting bugs. It's a opportunity for mentorship, and there's some amount of clairvoyance that senior engineers have on what a codebase should look like in the future that AI, at least today, cannot really replicate. And so we usually say, "Well, this is gonna do all the stuff that is not interesting for you to do, like just carefully reading the code for little mistakes, and you can focus on the bigger picture things as a human being."

   3. BFBrad Flora1:35

      W- um, and I understand you guys have several customers. You're reviewing a lot of

4. 1:40 – 2:10

   20,000 Bugs Found Every Week

   1. BFBrad Flora1:40

      lines of code. Like what, what are the numbers these days?

   2. DGDaksh Gupta1:42

      So in a given week, we usually review anywhere from five to eight million lines of code, and, uh, this is growing at a fairly high pace. This past week, we revealed twenty thousand bugs across the pull requests that Greptile reviewed for customers, so this is something that people like a lot. It's stuff that you-- that would've slipped through the cracks and would've caused problems later.

   3. BFBrad Flora2:00

      I understand that a lot of what you're doing is enabled by the language models that are out now. But, um, how do you think about like why now is the right time for Greptile, for a product

5. 2:10 – 3:20

   Why Now: LLMs and Code

   1. BFBrad Flora2:10

      that reviews code reviews?

   2. DGDaksh Gupta2:12

      Yeah, there's a couple of things. So my co-founders and I had started to experiment with large language models when DaVinci-two had just come out about, about three or four years ago. And we'd started to notice that there was a, um, clear use case in code because they understood code better than they understood maybe even la- natural language. And where when we see large language models today, they're, they still are better at reading code than they are at writing it. And to us, it's more interesting to take their current, uh, best capabilities and see what is the most value that they can create. We started by building code-based Q&A, so just loading up your codebase and asking questions about it, like, "How does XYZ work?" And we found that if you're able to teach a large language model to, um, to understand an entire codebase, the most useful thing you can do with that understanding is to find bugs in it and, and prevent code from going down. And I think the second major thing is we often think about this is a longer term trajectory, but software is not just toys, and like it was maybe fifty years ago. Like today, uh, we- our, our grid and our supply chains are controlled by software, and like, our world is only as resilient as our most-- our least resilient piece of software. And making software more resilient is something that only

6. 3:20 – 4:20

   From Code Q&A to Bug Detection

   1. DGDaksh Gupta3:20

      gets more important every year.

   2. BFBrad Flora3:21

      How did you guys get onto this idea? What led you here?

   3. DGDaksh Gupta3:24

      We started out, as I said, working on code-based context, and the reason for that was we were hacking on some projects in college and then one of our original team members decided to, uh, go and get a job and not continue on with us. And w- he had written our entire front end codebase, and we just didn't know how to work with it. And this was also something that we'd experienced at companies we'd worked at, which is codebases get large, and they're very difficult to work with. It's actually most of why the job of engineering is hard, is because codebases are big and complicated. And so we wanted to figure out if there was a way to take these enormous codebases and large language models that actually have limited context windows, find a way to teach them to completely understand the codebase and answer questions about it. So we focused on this problem for a long time, and then at some point, we realized that the most-- the, the problem of not knowing your codebase is temporary. Like, once you spend enough time at a company, you'll understand the codebase. But the problem that remains persistent is, uh, is, is finding bugs. And you can do that really well if you understand the codebase, which is what we had spent all this time working on.

7. 4:20 – 5:40

   What Developers Really Think of AI Tools

   1. BFBrad Flora4:20

      Okay, so you've built this product. You've got these users. You're reviewing all this code. What's, what's been surprising that you've learned along the way about AI and its relevance for developer and for developer tools?

   2. DGDaksh Gupta4:31

      I, I was surprised to learn how much variation there is in developers', um, propensity to adopt AI developer tools. There are people that are, uh, very, very smart developers that are completely against adopting any AI. And I think part of that, uh, might just be like discomfort around, um, around like AI maybe like affecting the nature of their job. Um, and then there's people at the very other end of the extreme, which are-- have very much strongly embraced, uh, AI as a tool. One of the things that we try to figure out when we're seeing if a customer is a good fit is just to see how high conviction they are around AI's ability to improve their business.

   3. BFBrad Flora5:06

      Mm-hmm. Do you, do you get lumped in with the AI coding tools, even though you're doing more code review instead of actual code gen?

   4. DGDaksh Gupta5:12

      This was a really big problem early on. And I think early on, uh, engineering leaders and, and engineers didn't have, uh, enough context to kind of separate these tools, and they all kind of looked like the same vague thing, which is AI in code. But as time has gone on and people have adopted some of these tools and they're using them, people have much higher fidelity in differentiating what different parts of their, uh, system can be augmented.

   5. BFBrad Flora5:33

      So a lot of the folks that watch these videos and tune in are people thinking about starting their own businesses. You were there not too long ago with the, like the early

8. 5:40 – 6:29

   Advice for Founders

   1. BFBrad Flora5:40

      days of Greptile. What advice, if any, would you have for someone that's thinking of setting out and doing a startup today?

   2. DGDaksh Gupta5:46

      I think, um, it's that there are many, many ways to get distracted. And, uh, often it-- no matter how much you are aware of the things that can distract you, you will-- or you're still not immune from distraction. You just really should focus on building products that your customers love and need and, and want, uh, and, and just not focus on anything else. I think there's a hundred other things that seem like they're things you should plausibly focus on, but they're just not as important as the core couple of things, and as long as you just do those, you can probably get pretty far.

   3. BFBrad Flora6:16

      If someone wanted to check out Greptile and learn more about it or even try it at their company, how would they do that?

   4. DGDaksh Gupta6:21

      So it's actually free to try. So you can go to greptile.com, and it takes less than a minute to sign up. You can just connect your GitHub and start using it.

   5. BFBrad Flora6:27

      Okay. Thanks so much for joining us.

   6. DGDaksh Gupta6:29

      Thanks so much.

Episode duration: 6:29