Latest / The Neon Show / Burning $8M on Open Source, Hitting $100M ARR and the $250 Billion Market | Harjot, CodeRabbit

Transcript

Burning $8M on Open Source, Hitting $100M ARR and the $250 Billion Market | Harjot, CodeRabbit · The Neon Show

0:00
SPEAKER_01: Companies are going to run as code.
0:01
SPEAKER_01: Everything is going to become a coding problem.
0:03
SPEAKER_01: If you can translate any kind of work to a coding problem, you will move 10x plastic.
0:07
SPEAKER_01: Code reviews were always a pain point.
0:09
SPEAKER_01: We have to show people how good AI is with code review.
0:12
SPEAKER_01: So he made it free for open source.
0:14
SPEAKER_01: We took a big hit.
0:14
SPEAKER_01: Now at that time, GPT-4 was$40 a million tokens.
0:19
SPEAKER_01: The recent round of 140 million at 1.5 billion.
0:22
SPEAKER_01: Coding space is always going to be like a lot of options.
0:24
SPEAKER_01: There will be cult followings.
0:26
SPEAKER_01: But code reviews is one of the spaces where you can actually make a good B2B business.
0:30
SPEAKER_01: We have been raising at a cadence of every 12 months.
0:32
SPEAKER_01: And in the next few months, we'll be crossing 100 million years.
0:36
SPEAKER_02: What roles are developers play then?
0:37
SPEAKER_02: Because if Arjot, welcome to the Neon Show.
0:44
SPEAKER_02: Excited to be doing this podcast.
0:46
SPEAKER_02: And I think this is very close to my heart since uh I've been an ML engineer in 2011.
0:53
SPEAKER_02: I was a top coder, ACM, ICPCM back then.
0:58
SPEAKER_02: Right.
0:58
SPEAKER_02: And the coding space has just exploded globally.
1:03
SPEAKER_02: Obviously, there is anthropic, open AI with Claude and Codex, right?
1:07
SPEAKER_02: But generally everybody wants to have a piece of the pie.
1:11
SPEAKER_02: So for our audience, if you can share from your lens, how does the coding space look like?
1:17
SPEAKER_01: Yeah, it's been a very interesting journey in the last few years for us.
1:21
SPEAKER_01: Like when we started back in 2023, the use of AI in code generation was just about taking off.
1:27
SPEAKER_01: Like if you recall, like GitHub Copilot was the first killer app on these models, even before ChatGPD predates that.
1:35
SPEAKER_01: And it was like a magical movement at that time.
1:37
SPEAKER_01: I was in my second startup, and you could just see like the world's gonna change.
1:41
SPEAKER_01: Like there will be a lot more code being written by AI than humans.
1:44
SPEAKER_01: And prior to that, we were, anyways, using Tab 9, like GPT-2, where are like they were not that good, but you could feel the difference.
1:50
SPEAKER_01: Copilar was just a different thing.
1:52
SPEAKER_01: And then cursor came along and so on.
1:54
SPEAKER_01: And then the another big inflection point happened when Claude Code came out.
1:59
SPEAKER_01: I mean, that's where you could see that the code editors are obsolete.
2:03
SPEAKER_01: Like you're you don't need to be navigating to the files manually.
2:07
SPEAKER_01: You can just now tell an agent in a chat what to do, and it goes and does it.
2:12
SPEAKER_01: And fast forward to now, we have had like great models like Opus 4.5 and now like continuing on that trajectory where these agents have become so capable running for longer and longer periods of time.
2:24
SPEAKER_01: Um, and like it's very clear to now, like what Dario was saying, that this is gonna get automated, it actually did, and faster than I could have anticipated, actually, uh, to a lot of extent.
2:35
SPEAKER_01: Yeah.
2:35
SPEAKER_02: So how did uh Code Rabbit come about and how did you fit so well into the tailwind?
2:41
SPEAKER_01: No, that's interesting.
2:42
SPEAKER_01: So this has been my third startup.
2:44
SPEAKER_01: So, you know, across all my startups, I've always looked for opportunities where um you look at like second-order effects.
2:52
SPEAKER_01: You cannot really create a wave, it's easier to ride a wave in the market.
2:55
SPEAKER_01: And when these AI agents were taking off, it was very clear that coding is going to be the execution layer for these agents, whether it's software development or any even business work, you're going to go through coding.
3:06
SPEAKER_01: Right.
3:07
SPEAKER_01: And as a result, uh, we could clearly see if you have ever on like a team or startups that code reviews were always a pain point.
3:16
SPEAKER_01: Even pre-Jennier, they were a pain point, by the way.
3:18
SPEAKER_01: Like, especially in fast-moving startups, like you would wait for like a few hours or a few days sometimes just to get an approval.
3:25
SPEAKER_01: Now, with coding agents taking off, it was very clear that it's gonna break and the choke point is gonna exist.
3:31
SPEAKER_01: So that's where we saw the opportunity.
3:32
SPEAKER_01: We built the product to solve like a more of a team-based pain point more than individuals' pain point, and got tremendous success in doing so.
3:42
SPEAKER_02: So um so was it always the vision to build a coding review product?
3:47
SPEAKER_01: That's right.
3:47
SPEAKER_01: So Code Rabbit from day one has been a code review solution.
3:51
SPEAKER_01: Now we always stayed away from code generation ourselves because that's a space where we feel um larger players um uh like OpenAI anthropic um are going to be more successful, even the open models now.
4:06
SPEAKER_01: And we also one of the things we saw early on, and we were right about that, that code generation will be commoditized.
4:11
SPEAKER_01: Okay.
4:11
SPEAKER_01: Like the nature of the space has been um that the developers will always pick up the tool they like to use.
4:18
SPEAKER_01: Like now it like earlier it was like Vim versus Emacs or VS Code.
4:22
SPEAKER_01: Now it's about the model choice, like some people like GPT-5, which is more pragmatic versus Opus, and then the choice of harness on top of it.
4:31
SPEAKER_01: Like they will use cloud code, open code, and then the surface they want to consume, whether it's in the editor still or in the CLI or desktop app.
4:39
SPEAKER_01: So coding space is always going to be like a lot of options and developers picking up what they like and there will be cult followings.
4:46
SPEAKER_01: But code reviews is one of the spaces where you can actually make a good B2B business, like actually selling to enterprises who have a pain point.
4:54
SPEAKER_01: Um, and and and very likely the most of the value we foresee is going to accrue in this segment of the market once that does settles.
5:01
SPEAKER_02: Understood.
5:03
SPEAKER_02: So so how big is if you have to put numbers, the coding, code generation market and the code review market?
5:09
SPEAKER_01: That's a great question.
5:10
SPEAKER_01: Code generation right now is probably several trillion dollars and growing, given how many people can now.
5:16
SPEAKER_01: In terms of potential revenue?
5:18
SPEAKER_01: Yeah, I mean, everything is going to become a coding problem.
5:20
SPEAKER_01: Like with the AI agents, if you can translate your pro any kind of work to a coding problem, you will move 10x faster.
5:26
SPEAKER_01: Whether it's making slide decks or running anything, even marketing.
5:30
SPEAKER_01: Like, I mean, companies are going to run as code, right?
5:32
SPEAKER_01: Very likely.
5:34
SPEAKER_01: So the in terms of and even like also the coding has become accessible to more people.
5:38
SPEAKER_01: Like earlier there was always a friction in learning languages and so on.
5:42
SPEAKER_01: Now you can just prompt, and anyone can build websites, internal apps, wipe code it, right?
5:47
SPEAKER_01: So the market has really grown and it and there's infinite appetite for software in the market, right?
5:53
SPEAKER_01: There's infinite appetite.
5:55
SPEAKER_01: So there's a massive, massive market.
5:57
SPEAKER_01: And secondly, if you look at code review, um, it is a subset of that market, right?
6:03
SPEAKER_01: Um, and if you look at the analogy, like if you remember Datadog used to provide observability, which meant which is almost like I don't know 25-30% of your AWS bill, like we kind of feel that code reviews and explainability, which is like how do I build trust in these outputs, is an equally valuable problem and roughly going to be like I don't know, 25-30% of your code generation market dam.
6:27
SPEAKER_02: You're saying that if if the code generation market becomes like a say a trillion dollar market in terms of revenue, not market cap is already there.
6:37
SPEAKER_02: But then uh code review could be$250 billion.
6:41
SPEAKER_01: Something like that.
6:42
SPEAKER_01: That's right.
6:42
SPEAKER_01: I mean, we're looking at that kind of a potential, if not more.
6:45
SPEAKER_01: Um, even in the past pre-Gen AI, the senior developers were spending 25-30% time doing code reviews, not writing code.
6:52
SPEAKER_01: Yeah.
6:52
SPEAKER_01: So that tax, code review tax has always been there even pre-Gen A, but now that is becoming even a bigger bottleneck.
6:58
SPEAKER_01: In fact, now people are spending more time reviewing code than writing the code, actually validating the changes.
7:04
SPEAKER_02: So let's assume, let's say, with Code Rabbit, if you can explain the product, uh the code is generated by Claude, and the code is automatically reviewed by Code Rabbit.
7:14
SPEAKER_01: So Yeah, it's kind of like what people call loop engineering, which is very interesting.
7:18
SPEAKER_01: Like when we started, humans were still looking at codabit feedback and then plumbing it back to and and doing tab completion on that.
7:25
SPEAKER_01: But ever since Cloud Code took off, like now we see Cloud Code just fighting with CodeRabbit in a loop and iterating over several commits to make sure CodeRabbit is happy and all the guardrails that it provides for the for the organization are it's compliant.
7:40
SPEAKER_01: Um so we see a lot of that pattern.
7:42
SPEAKER_01: Um yeah, so it's an adversarial review uh where it's having um a specialized harness with a very clean context around all the checklists that you want to enforce before you want to ship something to production.
7:54
SPEAKER_01: Got it.
7:56
SPEAKER_02: And uh so so what what roles does a developer play then?
8:00
SPEAKER_01: Because if you're developer, it's changed a lot even within our company.
8:06
SPEAKER_01: Like if you look at the entire frequency, taking a step back, like entire software development has been reordered.
8:13
SPEAKER_01: Earlier, your engineering resources were scarce, right?
8:18
SPEAKER_01: So you will have a lot of planning upfront, Jira, linear, so that you can regulate the work downstream.
8:24
SPEAKER_01: So you will have all the Scrum agile processes were built around making sure developers were always working on important things and prioritize accordingly, right?
8:33
SPEAKER_01: Now, what has happened?
8:35
SPEAKER_01: It's completely floodgates have opened.
8:37
SPEAKER_01: So you don't need to go to issue systems anymore.
8:40
SPEAKER_01: Even linear famously is saying issue systems are dead, which we clearly see.
8:43
SPEAKER_01: Now you can you are limited by imagination and we can go from idea to code, which means that not just developers, product managers, UX designers, and even tools like Datadog, Linear, Notion, uh, Post Hog, they're all becoming coding agent companies.
8:58
SPEAKER_01: And what that means for the developer is like they are not even like choosing what to build.
9:05
SPEAKER_01: Like most of these background agents, autonomous agents are now taking off.
9:09
SPEAKER_01: Every day, Datadog looks at their alerts and says, okay, I can actually fix this.
9:13
SPEAKER_01: And they're opening a pull request.
9:14
SPEAKER_01: Same is done with Sentry and security tools and so on.
9:17
SPEAKER_01: So what ends up in the new world is that developers are now becoming reviewers essentially.
9:22
SPEAKER_01: So the whole human judgment is about which tasks do I take to completion?
9:26
SPEAKER_01: Now the tasks are done by AI.
9:28
SPEAKER_01: 80% of the way AI does the work, that that last 20% is a human taste.
9:33
SPEAKER_01: Yeah.
9:33
SPEAKER_01: So there they bring the taste, they prioritize the work.
9:36
SPEAKER_01: So humans still have to decide out of this, I don't know, 100 pull requests in front of me opened by agents, which ones do I take to completion?
9:43
SPEAKER_01: Okay.
9:43
SPEAKER_01: Is it a security issue?
9:45
SPEAKER_01: Customer support tickets.
9:46
SPEAKER_01: So a lot of the planning is now happening in code, in the pull request.
9:50
SPEAKER_01: You don't have tickets anymore.
9:51
SPEAKER_01: You you just you only realize the importance of the task once it's already code.
9:56
SPEAKER_01: Yeah.
9:56
SPEAKER_01: So the role has changed a lot.
9:58
SPEAKER_01: I mean, um, humans are now just becoming reviewers and people who validate.
10:04
SPEAKER_02: And so so in this workflow, what what does CodeRabbit take away from the developer?
10:09
SPEAKER_01: It does a couple of things.
10:10
SPEAKER_01: First of all, even before a developer has to come and spend time on a code change, they can be sure that it's already meeting all the guidelines of the every organization have their own guardrails and everything.
10:21
SPEAKER_01: They want to make sure the code is maintainable, shippable, governable, like security posture is good.
10:26
SPEAKER_01: So the adversarial review loop happens even before a human comes in.
10:30
SPEAKER_01: So when cloud code agent opens, so they they trade.
10:34
SPEAKER_01: And so that's a big part that it takes away.
10:36
SPEAKER_01: And the second thing it does is it helps the developers triage and prioritize what to work on, right?
10:43
SPEAKER_01: So they are coming in and then they can say, okay, a customer impacting problem needs to go in first.
10:48
SPEAKER_01: And Code Rabbit helps them plan all that.
10:50
SPEAKER_01: And the third thing it does really well is explainability.
10:53
SPEAKER_01: So if I'm a reviewer, I'm coming into this pull requests, which are becoming larger and larger because the agents are now running longer with no one's reading the code.
11:01
SPEAKER_01: Authors are not reading the code, reviewers cannot really understand and piece together that puzzle.
11:06
SPEAKER_01: So what Code Rabbit has done is provides like higher order abstraction where line-by-line reviews done by agents, Code Rabbit, but uh humans come in and then they can understand the impact of those changes, the blast radius, the architectural impact in a more visual manner.
11:21
SPEAKER_01: It's all about trust.
11:22
SPEAKER_01: Now, if you can build trust in a pull request, you can then ship it safely and you're managing the risk.
11:26
SPEAKER_01: Otherwise, uh you basically are going to have a backlog of all these pull requests, and velocity is going to not change.
11:34
SPEAKER_01: And and which are the other companies in in the same space as Code Review?
11:38
SPEAKER_01: Yeah, so a lot of companies started around the time we started, and um we've been like clearly a market leader, but a lot of companies have been trying to uh like basically build a clone of Code Rabbit or become Code Rabbit like equivalent.
11:52
SPEAKER_01: So we have companies um like Kodo is another company which we see in a lot of enterprise deals, uh the CSB funded, then there is like uh Graptile from um YC ecosystem, uh, which is also like a smaller player.
12:07
SPEAKER_01: Um, then the model labs themselves uh offer some code review solutions which are not as effective, very expensive given the general purpose harnesses, they're not tuned for these kind of use cases.
12:18
SPEAKER_02: Got it.
12:19
SPEAKER_02: And and there are various benchmarks.
12:20
SPEAKER_02: So, for example, Graplit on one of their benchmarks said they are number one, but on independent benchmarks, Code Rabbit is number one.
12:28
SPEAKER_02: So, how how do those people believe these benchmarks?
12:30
SPEAKER_01: First of all, we ourselves don't believe any of these.
12:32
SPEAKER_01: Like I know there are a lot of benchmarks, but then directionally some of them, like first of all, the benchmarks coming from another vendor will always be biased.
12:39
SPEAKER_01: And that is one of the reasons we never did a benchmark because we we don't think we'll be able to do like a fair job in doing that benchmark.
12:47
SPEAKER_01: There are independent benchmarks uh which we follow, like there's a Martian benchmark we look at.
12:52
SPEAKER_01: But then again, like it then code review is such a nuanced space, like coming from outside, you're only looking at one aspect of it.
12:59
SPEAKER_01: Like, what's the North Star?
13:00
SPEAKER_01: Like, even the if you look at model labs, they themselves are now less focused on benchmark and optimizing for benchmarks versus like actually seeing the revenue and and what products people pick up.
13:11
SPEAKER_01: These are very different things, like the AI that humans can trust and actually find usable versus acing some benchmark and a test.
13:19
SPEAKER_01: Um yeah, and I I think the market is moving a little bit away from uh unless we like figure out like what the right benchmark is.
13:27
SPEAKER_01: Even we don't know how to really best express the best core reviews, like because you could look at things like accepted comments, but then the problem is like sometimes people will not choose to accept feedback because they don't want to bloat the existing PR, they'll do it later.
13:41
SPEAKER_01: Yeah, so it's very nuanced on how you measure efficacy of these systems.
13:48
SPEAKER_02: So, what is most interesting to me is uh uh in our conversation, you know, you have mentioned that though it's it's a very large market, at the top fewer and fewer companies are getting funded, let's say at above a billion dollar valuation.
14:03
SPEAKER_02: Uh, how many companies would be there?
14:05
SPEAKER_01: I mean, there have been a few.
14:06
SPEAKER_01: I mean, overall, if you look at the current fundraising market, it's a most weird market I've seen.
14:11
SPEAKER_01: Of course, AI is creating a lot of value, but a lot of the money is concentrating on anthropic, for instance, and OpenAI, some of these model labs, like some of the larger investors are completely indexed in anthropic right now.
14:24
SPEAKER_01: I don't know how whether it's a good thing or a bad given, how much open source is coming from behind and um potential to even like disrupt their revenue streams, right?
14:34
SPEAKER_01: There's going to be downward pressure.
14:35
SPEAKER_01: Um it's famously said it is almost like airline industry.
14:38
SPEAKER_01: It's cool, but it's very hard to make money where these guys are.
14:41
SPEAKER_01: Uh, on the other hand, um the startups uh have in general also struggling against cloud code.
14:49
SPEAKER_01: Like a lot of the companies got disrupted by cloud code, it's general purpose hardness.
14:53
SPEAKER_01: Cursor itself has to exit.
14:54
SPEAKER_01: I mean, they didn't they went to agents only after like Cloud Code had already started taking away their market.
15:00
SPEAKER_01: It was almost like a Yahoo versus kind of a Google movement for them, right?
15:03
SPEAKER_01: And we have seen perplexity and some of these startups um still still trying to find their own identity in this in this new world.
15:10
SPEAKER_01: Yeah, it's I mean, in general, like yeah, it's a kind of uncertain market for a lot of investors on where the value is gonna accrue.
15:16
SPEAKER_01: Um but if you ask me, I mean, I would say a lot of the value is going to be in the in these applications, the surfaces, because the models are gonna get commoditized.
15:25
SPEAKER_01: I mean, models are such that it's so easy to copy and replicate them, right?
15:30
SPEAKER_01: I mean, that's what we have seen with all this like GLM and Kimi, they're becoming as good as, especially with model routing, you can actually squeeze as much performance as fable or something.
15:39
SPEAKER_01: Yeah, right.
15:40
SPEAKER_01: Um, so that trend is gonna continue.
15:42
SPEAKER_01: The harnesses itself, which are running locally, they are gonna be open source.
15:46
SPEAKER_01: There is going to be someone who's gonna emerge as, I don't know, Kubernetes of coding harnesses.
15:50
SPEAKER_01: Like I don't know whether it's open code or someone else, but it's inevitable that's gonna happen.
15:55
SPEAKER_01: So, where's the value is gonna accrue?
15:57
SPEAKER_01: It's going to be in products like Code Rabbit, which are actually solving um like a team-based train point.
16:02
SPEAKER_01: That surface is very complex for everyone to access.
16:05
SPEAKER_01: Like, for example, we have the most GitHub in installs of any AI product, uh, which is a very hard surface to go and crack open.
16:12
SPEAKER_02: So, so congrats on raising the recent round of 140 million at 1.5 billion.
16:17
SPEAKER_01: That's right, yeah.
16:18
SPEAKER_01: So we have been like raising at a cadence of every 12 months.
16:22
SPEAKER_01: I mean, this is the first time I've seen things move on such a compressed time frame.
16:26
SPEAKER_01: Normally, companies will take a little bit longer to go to Series A, B, C.
16:29
SPEAKER_01: For us, it's been kind of a hypergrowth trajectory and it's only accelerating, actually, which is very interesting.
16:35
SPEAKER_01: Even right now, every week we have a new maxima with the number of daily active users or new signups.
16:41
SPEAKER_01: Um, so as these agents become more affordable, they're taking off, getting more commoditized, the more you need code reviews.
16:47
SPEAKER_01: And what are the like today's daily active users?
16:49
SPEAKER_01: It's several hundred thousand actually, and all paid, like like I would say 15 to 20% users will be open source because we give product uh free for open source users, which help us improve our hardness a lot.
17:01
SPEAKER_01: Um, then I would say 80% will be like the paid users out of that.
17:04
SPEAKER_01: And and these are individuals or mostly across enterprises?
17:08
SPEAKER_01: Yeah, we don't market individuals, but still a large number of individual developers buy Code Rabbit because they really love the adversarial, the loop engineering, the feedback loops.
17:18
SPEAKER_01: Uh, the product is primarily geared and optimized towards teams and um, like the authors, the review of the team setting.
17:25
SPEAKER_01: The development teams buy us as a whole.
17:27
SPEAKER_01: So they each time we do a trial, there's always a survey where 80 to 85% developers uh say that they cannot go back to world without co-driver.
17:35
SPEAKER_01: So a very team-based product, uh, which also makes it very sticky.
17:38
SPEAKER_01: Like the churn is super low in this kind of a business.
17:40
SPEAKER_01: Once you're part of a workflow, you are you're there to stay.
17:44
SPEAKER_02: And and how do you price the product like which has worked so well?
17:47
SPEAKER_01: Yeah, it is a very big challenge.
17:49
SPEAKER_01: In a way, like the market um wants to see per seat pricing, but you know, like the usage-based pricing is where a lot of the revenues for other companies are coming from.
17:59
SPEAKER_01: So we price it by seed to a large extent.
18:01
SPEAKER_01: We also have a usage model on top when you are like becoming a power user and you are hitting the limits.
18:06
SPEAKER_01: Uh, we have some reasonable rate limits, but beyond that, we charge by the usage.
18:11
SPEAKER_01: Most of the revenue does come from seed-based pricing because we have very liberal and reasonable rate limits.
18:16
SPEAKER_01: Most people don't even hit those.
18:18
SPEAKER_01: Um, by the way, which is a very hard engineering problem.
18:21
SPEAKER_02: Yeah, because I think I saw on your website there are two pricing models:$24 and$48.
18:25
SPEAKER_02: That's right.
18:26
SPEAKER_01: Yeah, so there's a pro plan and there's a pro plus plan, and the difference is in the features and the rate limits you get.
18:32
SPEAKER_01: Um, but in general, like it's very predictable per se pricing, by the way, which is very hard engineering challenge, to be honest.
18:37
SPEAKER_01: Like using an ensemble of model, for example, we don't let users use pick a model from a top-down.
18:43
SPEAKER_01: Like, I mean, like if you look at every other tool, like cursor and all or cloud code, you are always picking a model.
18:48
SPEAKER_01: So we've always been an ensemble of models.
18:51
SPEAKER_01: We have always been very efficient with sub-agents, context engineering on bringing the right tokens in front of the model, especially the reasoning models, which can be very expensive.
19:00
SPEAKER_01: Um, and that has led us to like being kind of like Amazon Prime, like, yes, everyone wants free shipping, but practically it's very hard to offer that.
19:07
SPEAKER_01: But how much is your margin uh when you're offering the C pricing?
19:11
SPEAKER_01: Yeah, it's improving actually.
19:12
SPEAKER_01: So uh in fact, it's very interesting.
19:15
SPEAKER_01: Like as the time goes by, like as our own harness becomes more and more efficient, uh, we are already always like like on a trajectory to be like 50 to 80 percent margin business already.
19:28
SPEAKER_01: Like uh and and part of it is like also the downward trend in the model pricing.
19:32
SPEAKER_01: Yeah, like the models themselves, uh thanks to open source, uh, even the premium models with OpenAI Anthropic are overall getting cheaper.
19:40
SPEAKER_01: Even OpenAI last week did a big price drop for some of the models which we use.
19:45
SPEAKER_01: And secondly, it's it's the context.
19:48
SPEAKER_01: Like we are also getting really good at uh sandboxing the infrastructure optimizations that we do around that.
19:54
SPEAKER_01: And um even even using a lot of code graphs, for example, like we don't have to always naively navigate the code to figure out the blast radius.
20:02
SPEAKER_01: So we do a lot of semantic analysis to very efficiently bring the right context in.
20:09
SPEAKER_02: So you would say roughly you would make 50% gross margins?
20:13
SPEAKER_01: That's right, we are on that trajectory already.
20:15
SPEAKER_01: So it's not negative.
20:16
SPEAKER_01: I mean, yeah, but it's like becoming better and better over time very quickly.
20:20
SPEAKER_02: But most of these companies, I think investors are worried right now, if OpenAI and Anthropic go public, they are spending trillions of dollars on infrastructure every year and are on negative gross margins.
20:31
SPEAKER_02: So the entire industry might not not collapse.
20:34
SPEAKER_01: No, it's not gonna happen.
20:36
SPEAKER_01: Like it's just compute end of the day, right?
20:37
SPEAKER_01: I mean, and VDI is doing a great job in making more efficient uh GPUs and those clusters, right?
20:44
SPEAKER_01: At the same time, the models themselves are getting smaller but smarter, right?
20:48
SPEAKER_01: And the thirdly, the inference, the stack, the caching, um, very there's a lot of low-hanging fruit in how you bash, how you cache, and how you serve models at scale.
20:59
SPEAKER_01: Um, and this is where like some of the open AI cost changes are coming from.
21:03
SPEAKER_01: I mean, they've been just optimizing their inference stack, right?
21:06
SPEAKER_01: So I I think when open AI says they're uh like too cheap to meter, it's actually true.
21:13
SPEAKER_01: I mean, the way we are heading is like um this is gonna be like like pennies at at some point.
21:20
SPEAKER_01: Yeah.
21:21
SPEAKER_02: So how will they make the margins to to build?
21:24
SPEAKER_01: You'll just use more AI, right?
21:26
SPEAKER_01: I mean, you'll use more subagents, you will use more tokens per minute than you are doing right now.
21:30
SPEAKER_01: Right now, in very early stages, like if the model prices go down, then you're just gonna spin up more subagents, spend more tokens, do more stuff, right?
21:38
SPEAKER_01: More um test and compute, right?
21:40
SPEAKER_01: Um very clearly.
21:41
SPEAKER_01: So even we have seen that, like as the model prices go down, our margins have trained stayed constant because we just end up using more AI, adding more sub agents, being more thorough.
21:50
SPEAKER_01: So the product improved from last year from I don't know to 35, 30 to 35% acceptance rates to over 65-70% now.
21:59
SPEAKER_01: And that's Thanks to cheaper models leading to more thorough reviews.
22:03
SPEAKER_02: So, how do you and Gur know each other?
22:05
SPEAKER_02: Like, and what's your history of working together?
22:07
SPEAKER_01: Yeah, I mean, we're related actually.
22:08
SPEAKER_01: So, Gur is my co-founder, by the way.
22:10
SPEAKER_01: So, he started Code Rabbit.
22:12
SPEAKER_01: So, Code Rabbit is a very interesting history.
22:15
SPEAKER_01: Um, so it started it as a kind of a um weekend project in my second startup.
22:21
SPEAKER_01: So, we had like 15 developers who were starting to use GitHub Copilot at that time, and code reviews were very clearly broken.
22:28
SPEAKER_01: And especially it was a remote team, so they were in Eastern Europe or something, right?
22:32
SPEAKER_01: And that's where like I and Gore had recently got his green card.
22:36
SPEAKER_01: I mean, that's always like a big challenge with uh a lot of Indian founders is you're on H1B and you're entrepreneurial, but you can't do anything.
22:43
SPEAKER_00: Yeah.
22:43
SPEAKER_01: So as soon as he got his green card, I mean he was looking to start something in the AI space, and this is something which was close to his heart.
22:49
SPEAKER_01: He's been an engineering leader who understood uh the workflow really, really well and DevTooling.
22:55
SPEAKER_01: So he took that early prototype, actually productized it, um, and made it so delightful.
22:59
SPEAKER_01: Where if you look at most AI products were getting rejected at that time in these surfaces, like Microsoft putting Copilot everywhere, didn't work out that well, if you remember, right?
23:09
SPEAKER_01: So he managed to bring a very delightful experience, great DXUX, into a very nuanced workflow to the audience, which is very opinionated.
23:18
SPEAKER_01: Yeah.
23:18
SPEAKER_01: And people loved it.
23:20
SPEAKER_01: And the second thing he did really well was he made it free for open source and it just took off.
23:26
SPEAKER_01: And it was around that time that I realized in my second startup that that one's not gonna go anywhere.
23:31
SPEAKER_01: Um hard to get funding in infrastructure space at that time.
23:34
SPEAKER_02: How much did you raise that startup?
23:35
SPEAKER_01: That was like six and a half million raised, uh, only one round of funding.
23:40
SPEAKER_01: And this was pre-Gen AI, right?
23:42
SPEAKER_01: When Gen AI happened, uh, all the VCs were focused on just AI sector.
23:47
SPEAKER_01: Like it was like if you're not doing AI, you're gonna die.
23:50
SPEAKER_01: Like a slow death.
23:52
SPEAKER_03: Yeah.
23:52
SPEAKER_01: Right?
23:52
SPEAKER_01: So it was very clear.
23:53
SPEAKER_01: So I basically ended up uh talking to Gore and he agreed to acquire code my second startup, Lux Ninja.
24:00
SPEAKER_01: And that's how I ended up.
24:01
SPEAKER_01: So the company was already like a million ARR, uh, pretty much bootstrapped um when I joined.
24:08
SPEAKER_01: And right after I joined, then we did the series A round of funding, which was led by CRV at that time.
24:12
SPEAKER_02: And how did the next round of funding happen?
24:14
SPEAKER_02: Like what milestones?
24:15
SPEAKER_01: It was very quick.
24:16
SPEAKER_01: So we went from one to over 10, 11 million ARR in less than 12 months.
24:21
SPEAKER_01: Okay.
24:22
SPEAKER_01: Right.
24:22
SPEAKER_01: And that was a series B milestone for us.
24:24
SPEAKER_01: How much did you raise then?
24:26
SPEAKER_01: That one was a$60 million round done by scale ventures.
24:29
SPEAKER_01: So Andy Vetus is on the board.
24:31
SPEAKER_01: He loves the product, he actually pays for the product.
24:33
SPEAKER_01: Okay.
24:33
SPEAKER_01: So he had been using the product uh had been visiting our office, and right then we said we want to do fundraising, we got a term sheet.
24:39
SPEAKER_01: Yeah.
24:39
SPEAKER_01: Now it was a very quick fundraise.
24:41
SPEAKER_01: Same thing happened now.
24:42
SPEAKER_01: CDC was also very quick, even though it's a very uh interesting time to raise money given how much uncertainty is like on the coding space and the workflow in general.
24:52
SPEAKER_01: Uh, it was a very quick fundraise, even like this 143 million raised.
24:56
SPEAKER_02: So you also mentioned that uh the uncertainty so much that people don't know whether GitHub will exist today, tomorrow or not.
25:03
SPEAKER_01: I had to fail these questions, man.
25:05
SPEAKER_01: Like, which was so funny, right?
25:07
SPEAKER_01: The reason GitHub is under so much distress is a sign of things to come.
25:10
SPEAKER_01: I mean, people are opening so many pull requests, and and that's just an early sign on how important GitHub is going to be.
25:16
SPEAKER_01: It's a center of gravity.
25:18
SPEAKER_01: I mean, infrastructure issues are just temporary.
25:20
SPEAKER_01: They're gonna go work around it.
25:22
SPEAKER_01: I mean, it's just that we don't have they don't have capacity which is hitting them right now.
25:26
SPEAKER_01: Yeah.
25:27
SPEAKER_01: Um, so I don't think GitHub can even go away.
25:30
SPEAKER_01: First of all, that's the only choke point which is now remaining.
25:33
SPEAKER_01: See, the upstream choke point that used to be linear Jira, that's not there anymore.
25:36
SPEAKER_01: Yeah.
25:37
SPEAKER_01: People are bypassing issue systems.
25:38
SPEAKER_01: Yeah, the entire planning is happening in the code.
25:40
SPEAKER_01: It is.
25:41
SPEAKER_01: Like now it's because the agents are really good with code, that's the only thing they know.
25:44
SPEAKER_01: They cannot you cannot just stop them to submit a plan which you collaborate.
25:48
SPEAKER_01: I mean, that's a more of a vitamin use case.
25:50
SPEAKER_01: Like it's there's no choke point where you need an approval to write code.
25:54
SPEAKER_01: So, what is happening is tools I created and all are going straight to code.
25:59
SPEAKER_01: Humans are going straight to code from idea to PR.
26:02
SPEAKER_01: They are completely bypassing linear and issue systems.
26:05
SPEAKER_01: And all that, all those ideas as code are hitting GitHub directly.
26:11
SPEAKER_01: Now, more a lot of them are gonna get closed.
26:14
SPEAKER_01: But in the end, GitHub has to deal with all that influx.
26:17
SPEAKER_01: Uh and that's the only choke point.
26:19
SPEAKER_01: You cannot like have a world where you will ship all this stuff without looking at it once.
26:24
SPEAKER_01: Right?
26:25
SPEAKER_01: And that workflow is get end of the day.
26:27
SPEAKER_01: That is a place where you're gonna build a preview environment, you're gonna test the code, verify, bring the taste in, and then ship.
26:35
SPEAKER_02: So so then the entire market has put so many players under pressure.
26:39
SPEAKER_02: Do you think Figma will exist?
26:42
SPEAKER_01: No, that's a great point.
26:44
SPEAKER_01: I mean, a lot of people are not going directly to the UI.
26:48
SPEAKER_01: So a lot of these tools in the past were built around humans going to these tools and operating in them.
26:53
SPEAKER_01: Like, look at even dashboarding tools.
26:56
SPEAKER_01: We use Datadog, but I don't log into that UI anymore.
26:59
SPEAKER_01: I ask an agent which will call that API and give me exactly what I need.
27:04
SPEAKER_01: Same thing with Figma.
27:05
SPEAKER_01: Like, I mean, people agents are getting good with design as well.
27:09
SPEAKER_01: So it will exist, but the interface is gonna look very agentic.
27:13
SPEAKER_01: So a lot of code getting generated to do design, just like how we use code generation to do code review.
27:19
SPEAKER_01: But but Figma doesn't operate on code, right?
27:21
SPEAKER_02: It operates on human taste and design and social perfecting things.
27:25
SPEAKER_01: Yeah, I mean it's gonna exist.
27:26
SPEAKER_01: That's where the taste is still there.
27:28
SPEAKER_01: But whether the product is gonna look the same, that I don't know.
27:30
SPEAKER_01: How much the product has to evolve for agents, but the humans still have to bring in some original taste.
27:35
SPEAKER_01: Right now, whether it happens now in Cloud Code or happens in Figma is the question now.
27:40
SPEAKER_01: I have like the design has to happen somewhere.
27:42
SPEAKER_01: But Cloud Code itself is now in encroaching Figma's territory, if you have seen their trajectory and where it's going.
27:47
SPEAKER_01: And and the interface has changed basically now.
27:49
SPEAKER_01: Humans want to talk to an agent to describe the design, which they can iterate on versus going and clicking manually.
27:56
SPEAKER_02: But what do you think?
27:57
SPEAKER_02: Because let's say a designer would make things pixel perfect.
28:04
SPEAKER_02: But but can a code uh but design has never been a mathematical problem.
28:09
SPEAKER_01: No, no, the humans still have to bring the taste.
28:11
SPEAKER_01: So but but the question is like there's some tasks where agents can run autonomously for several hours, but design is something you have to tactically still drive.
28:18
SPEAKER_01: Right?
28:18
SPEAKER_01: You are having this interactive session with an agent back and forth until you get exactly what you want.
28:24
SPEAKER_01: So you really have to be very, you have to describe your destination.
28:27
SPEAKER_01: So some of the tasks have become like a self-driving car.
28:30
SPEAKER_01: You just punch in your destination, it takes you there.
28:32
SPEAKER_01: Things like refactors or migrations, that's where a lot of these background agents are also very successful.
28:38
SPEAKER_01: You can run a goal command, but you cannot run a goal command on design.
28:42
SPEAKER_01: Like you can ask the agent to like iterate to the point of diminishing returns.
28:46
SPEAKER_01: If you have like a good design system, it can make sure everything is consistent, can translate that to uh more applied.
28:52
SPEAKER_01: But if you're already talking about coming up with uh the new taste and the new design, you still have to drive the agents tactically.
28:59
SPEAKER_01: But one thing is very clear agents are everywhere, whether you're tactically driving or they're driving on their on their own for a longer period.
29:06
SPEAKER_02: Understood.
29:07
SPEAKER_02: So design is a space that is changing very, very rapidly.
29:10
SPEAKER_01: Right, right.
29:11
SPEAKER_01: And and we love uh we have like what we are doing is like now even converting a lot of our marketing to codes.
29:17
SPEAKER_01: So we have a repository where we have a common design system, all our slide decks, sales, marketing decks are done by AI.
29:24
SPEAKER_01: So our designer basically defines the design systems and the what colors, branding, everything.
29:31
SPEAKER_01: Like logos, all the templates are there.
29:34
SPEAKER_01: But then the agents are able to mechanize that foundational work into any kind of derivative content we need to create in the market.
29:41
SPEAKER_01: And we are moving 10x faster with that.
29:43
SPEAKER_01: Even Google ads are done through code in at Code Rabbit.
29:46
SPEAKER_01: So we have a GitOps method, like we actually open a pull request to deploy an ad.
29:49
SPEAKER_01: Okay.
29:50
SPEAKER_01: Right?
29:50
SPEAKER_01: It's all automated with agents right now.
29:52
SPEAKER_01: So how many team members do you have today?
29:54
SPEAKER_01: In terms of our company.
29:56
SPEAKER_01: Globally, we have like 250 to 275, I guess, now.
30:00
SPEAKER_01: Um, most of them are interestingly sales and marketing.
30:02
SPEAKER_01: Like you don't really need that many engineers.
30:05
SPEAKER_01: So even within the company, the bottleneck shifted.
30:07
SPEAKER_01: Like earlier, you will need a little bit larger engineering team at this stage.
30:10
SPEAKER_01: Now, with the same engineers over here, you are able to keep a large uh sales and marketing GTM team very busy.
30:17
SPEAKER_01: In fact, our bottleneck is not product creation anymore.
30:21
SPEAKER_01: We are creating products before faster than I would say we are able to talk about those products.
30:27
SPEAKER_01: The bottleneck is human attention.
30:30
SPEAKER_01: Like, how do we go to market and put the stake in the ground and um capture the mind share around new concepts that we are releasing?
30:41
SPEAKER_02: This is very interesting, right?
30:42
SPEAKER_02: Because if you're saying how many engineers do you have today?
30:45
SPEAKER_01: We have roughly 35 to 40, so that's it.
30:49
SPEAKER_01: Okay.
30:50
SPEAKER_01: And it's not like adding more engineers, it's gonna just create more chaos at this time.
30:54
SPEAKER_01: Because now we are just, as I said, limited by imagination.
30:57
SPEAKER_01: These agents are really uh double, tripled our velocity in the last three months.
31:02
SPEAKER_01: Right?
31:02
SPEAKER_01: Our biggest bottleneck is like now, as I said, talking about these products and the release velocity has uh become almost like a daily thing now.
31:11
SPEAKER_01: So, how do you do more like newsroom style marketing?
31:13
SPEAKER_01: So that those are the things we are now figuring out in the company.
31:18
SPEAKER_02: And um in in your journey, what do you think in Cold Rabbit were the things that went right?
31:26
SPEAKER_02: Or decisions that you were the founders uh made right that took you to where you are?
31:31
SPEAKER_01: Lot of decisions went right as well.
31:34
SPEAKER_01: Some of them maybe didn't pan out to be as great, but I would say mostly decisions have been right.
31:39
SPEAKER_01: Um even like doing things even before industry realized they were the right things to do.
31:44
SPEAKER_01: For example, if you can share stories and exactly.
31:46
SPEAKER_01: Yeah, even on the architecture, like we were one of the very first companies uh doing sandboxes.
31:51
SPEAKER_01: Like now the sandboxes are a common thing we were doing back in GP days of GPD 3, 5 and GPD-4.
31:57
SPEAKER_01: So while Cursor was doing, for instance, embeddings and similarity search for tab completion, we were doing what Cloud Code was doing back in 2023, doing all this code generation to do code review, all the CLI commands.
32:10
SPEAKER_01: So we were like one of the first companies to do sandboxes and all these like shell scripts to navigate the code with ripcrap back in 2023.
32:18
SPEAKER_01: And GPT-4 actually was very expensive back then.
32:20
SPEAKER_01: So GPT3.
32:28
SPEAKER_01: Which now people are talking about.
32:29
SPEAKER_01: So we have always been an ensemble of models.
32:32
SPEAKER_01: We never did MCPs, like our marketing team would come and pang me, like why are we not talking about MCPs?
32:37
SPEAKER_01: I'm saying no, the CLI commands is the right way, the training data is already there.
32:41
SPEAKER_01: You don't want to waste the tokens.
32:42
SPEAKER_01: How are agents talking to each other without MCP?
32:45
SPEAKER_01: No, we uh we are basically doing a lot of code generation live, like even let's say to open a Jira ticket or linear ticket in Code Rabbit, it writes a curl command to talk to that API.
32:56
SPEAKER_01: So a lot of the work happened in sandboxing and securing those sandboxes, like with proxies, credential injection.
33:01
SPEAKER_01: So a lot of these things became mainstream in 2026, I would say now.
33:05
SPEAKER_01: They're becoming mainstream now, which we were doing three years prior.
33:08
SPEAKER_01: And those were the right patterns, and these patterns emerged, right?
33:12
SPEAKER_01: And the other thing we did really well has been around distribution.
33:16
SPEAKER_01: Like as I mentioned, like my good my co-founder, not just build the product.
33:19
SPEAKER_01: I mean, of course, there's a lot of things in the DX we did got, right?
33:22
SPEAKER_01: But then the question was, how did you distribute these things?
33:24
SPEAKER_01: At that time when AI was so new, like a lot of people had not even used Chat GPT when we started.
33:29
SPEAKER_01: So their first time they were using an advanced AI was through chat with Code Rabbit.
33:34
SPEAKER_01: Right?
33:34
SPEAKER_01: The idea there was like we have to show people how good AI is with Code Review, and the best way to do was distribute it for free.
33:41
SPEAKER_01: So he made it free for open source.
33:43
SPEAKER_01: We took a big hit.
33:43
SPEAKER_01: Now, at that time I remember GPT-4 was$40 a million tokens.
33:48
SPEAKER_01: We still made it free and we were bootstrapped.
33:51
SPEAKER_01: So, how much money did you and Guru put together before you reached?
33:54
SPEAKER_01: I guess we would have collectively put in like a million and a half.
33:58
SPEAKER_02: Wow.
33:58
SPEAKER_01: Right.
33:59
SPEAKER_02: Is Guru also a second time founder?
34:01
SPEAKER_01: No, it was his first time, but he had some savings to do his own startup, uh, which he invested and wasn't taking salary for a while.
34:08
SPEAKER_03: Yeah.
34:08
SPEAKER_01: And then I had also put in some money then.
34:11
SPEAKER_01: Um there was Flex Capital as well, which came in right after um uh after me.
34:17
SPEAKER_01: Um so they put in like I don't know 750k at that time.
34:20
SPEAKER_01: So roughly 1.5 to 2 million is what we had.
34:23
SPEAKER_02: It was quite amazing that you know uh you took 750k from a fund uh and it it turned out so well for everyone.
34:32
SPEAKER_01: It did, yeah.
34:33
SPEAKER_01: We didn't raise a lot of money.
34:34
SPEAKER_01: Uh we took some risks, as I said, like even decision around making it free for open source on our own money, which is a risky thing, but then we were able to then prove to the world get the love, and that started the whole flywheel.
34:47
SPEAKER_01: And even like I was looking at our um overall spend.
34:51
SPEAKER_01: We have spent like seven or eight million on AI costs on supporting open source over the last three years.
34:59
SPEAKER_02: And if you can share uh current revenue scale of the company, any ballparks around it, and how is it how fast it's growing monthly?
35:06
SPEAKER_01: It's growing like roughly 15 to 20 percent each month.
35:11
SPEAKER_01: Yeah, I mean cheap keeps changing.
35:13
SPEAKER_01: Last time I checked, we were like north of 60 to 65 million ARR.
35:19
SPEAKER_02: Wow.
35:20
SPEAKER_02: And this is without credits pure revenue?
35:23
SPEAKER_01: Yeah, that's right.
35:23
SPEAKER_01: It's pure ARR, which is seat-based pricing.
35:26
SPEAKER_01: Or if I count overages, then it's even higher because some people do use it, but we don't count that as ARR.
35:31
SPEAKER_01: I know a lot of companies take their accounting.
35:34
SPEAKER_01: We don't do that.
35:34
SPEAKER_01: Like if for us it's actually seat-based pricing, recurring revenue on charge B or something.
35:38
SPEAKER_01: So we're very close to 100 number.
35:41
SPEAKER_01: Yes, I mean it's just around the corner.
35:43
SPEAKER_01: So we're very confident in the next few months we'll be crossing 100 million years, especially with acceleration we are seeing.
35:49
SPEAKER_01: We're actually now also seeing a little bit of an acceleration with demand.
35:53
SPEAKER_01: And what's the biggest GTM right now?
35:55
SPEAKER_01: It's all word of mouth, PLG.
35:58
SPEAKER_01: How is that possible?
35:58
SPEAKER_01: Like as I mentioned, like that whole open source thing was the key.
36:02
SPEAKER_01: I mean, because a lot of developers could see us in action in the projects they love, and we are in very popular projects.
36:09
SPEAKER_01: And that brings us, and they go and then bring us into their workplace.
36:13
SPEAKER_01: Um, and also like our existing customers really um strongly refer us to other customers.
36:18
SPEAKER_01: So if you and we do a survey on where you heard about us, number one is open source, second is just like people talking about us, right?
36:25
SPEAKER_02: And then paid ads and everything's like nobody's so love to know about your background uh and your co-founder background.
36:34
SPEAKER_02: Yeah, happy to talk about that.
36:36
SPEAKER_02: So which part of India you grew up in?
36:38
SPEAKER_02: How did you come to the US?
36:40
SPEAKER_02: So very entrepreneur family.
36:42
SPEAKER_02: Who is your brother-in law?
36:45
SPEAKER_01: Yeah, yeah.
36:46
SPEAKER_01: So we have always like so all of us belong to Punjab.
36:49
SPEAKER_01: So the father-law of an entrepreneur?
36:51
SPEAKER_01: No, he's he's entrepreneurially, but he never actually did us business.
36:54
SPEAKER_01: But he's always been like that growing up encouraged us to take all these risks in life, right?
37:00
SPEAKER_01: So it was never about like go and do a job in a bigger company or something.
37:03
SPEAKER_01: It was always about go out um and and build.
37:07
SPEAKER_01: I mean, because he's a builder, I mean, he's very uh tech forward looking.
37:10
SPEAKER_01: I mean, he knows what does he do?
37:12
SPEAKER_01: Well, he's he's an engineer, so he was uh in the government, he was a chief engineer and one of the people who also brought a lot of IT into the government at that time.
37:21
SPEAKER_01: Um like automatic digitized a lot.
37:23
SPEAKER_01: I remember like he knew more about computers than I do.
37:27
SPEAKER_01: I mean, even today, I mean he will go and show me the apps that even I have no idea, man.
37:31
SPEAKER_01: Like, I mean, he he's he he's really, really forward-looking with these kind of things.
37:35
SPEAKER_01: Uh, Gore has always also been like uh my co-founder, uh, grew up in Punjab, Bambrisar.
37:42
SPEAKER_01: And for him, the biggest constraint had always been like the green card.
37:45
SPEAKER_01: I mean, he I remember like the first time he tried to do something was in the IoT space.
37:50
SPEAKER_01: I was just mentioning you, right?
37:51
SPEAKER_01: When Sem Sarah started cold chain, uh but then all these visa issues didn't ban out.
37:56
SPEAKER_01: It was just so hard, like chicken and egg problem, right?
37:58
SPEAKER_01: And very luckily, like right when AI was getting started, I mean, he uh got his green card and quit right away to start Code Rabbit, right?
38:08
SPEAKER_01: Yeah, um, so very interesting luck in terms of timing, green card situation, and so on.
38:13
SPEAKER_01: Otherwise, we would not have been doing this, and because I was busy with my second startup.
38:17
SPEAKER_01: For me, I have like came to the US um back in 2008.
38:22
SPEAKER_01: And you grew up in I grew up in India and I came after uh doing my first job briefly for um I don't know two and a half years.
38:30
SPEAKER_02: Like in Punjab, which cities you lived in, which college you went to?
38:34
SPEAKER_01: Yeah, I grew up in different parts of Punjab, like uh Patiala is my schooling YPS Patiala.
38:40
SPEAKER_01: And then I grew up um also been in Chandigar, Jilanda, fewer, few other cities.
38:45
SPEAKER_01: Um, then I worked in Noida for two and a half years in Al Qadel, Lucent.
38:50
SPEAKER_01: That was my first big company exposure.
38:52
SPEAKER_01: Like, learned a lot there actually.
38:54
SPEAKER_01: And in fact, a lot of our my startup ideas came from that.
38:56
SPEAKER_01: I mean, like some of the tooling we were doing in telecom back then was not even something this web um scale computing people were doing.
39:04
SPEAKER_01: Like telecom was really technically advanced back in 2005-6.
39:08
SPEAKER_01: So that's when I first had an idea that okay, I mean, I want to do a startup in a tech sector.
39:12
SPEAKER_01: Like in back in India, 2005 and 6, there were not many VCs who would understand um some of this like um distributed systems and systems startups.
39:22
SPEAKER_01: Um, and that was the reason I came to the US when I um did my statement of purpose for that mission.
39:28
SPEAKER_01: I mentioned that I want to do a startup.
39:31
SPEAKER_01: And uh I got into a grad school at University of Pennsylvania.
39:34
SPEAKER_01: It's an IV league, it's in the East Coast, really good school.
39:36
SPEAKER_01: And you opted for PhD.
39:38
SPEAKER_01: Yes, so I came in, um, but the purpose there was never to actually I mean the main idea was to like do some research and see if that can be commercialized.
39:47
SPEAKER_01: So my first startup was exactly that.
39:49
SPEAKER_01: So it was a pen spin out.
39:52
SPEAKER_01: Um I never finished my PhD, but we licensed the technology from Penn and started a company.
39:58
SPEAKER_01: Um it started in Philadelphia, uh, but then that was again not a great place to do uh tech startup.
40:05
SPEAKER_01: I mean, most of the startups there were like FinTech or consumer tech startups.
40:09
SPEAKER_01: Uh they would VCs wouldn't understand in New York, in Boston, like what we were doing.
40:13
SPEAKER_01: So we I came to the Bay Area in 2015.
40:16
SPEAKER_01: And this place is how long?
40:18
SPEAKER_02: Like uh you came to US in 2008.
40:20
SPEAKER_02: Yeah.
40:21
SPEAKER_02: 2002, you I assume you would have quit your PhD.
40:24
SPEAKER_01: Right.
40:24
SPEAKER_01: So I came in 2008, so I quit my PhD in 2014, 2013 or 14.
40:30
SPEAKER_01: And I had been like building this project that I wanted to commercialize.
40:34
SPEAKER_01: So it was a hard to build.
40:36
SPEAKER_01: Um, in fact, it was a precursor to a lot of these eBPF startups, which in observability.
40:41
SPEAKER_01: So they were like, what we were doing um at NetSil was my first startup's name.
40:45
SPEAKER_01: It was like listen spelled backwards.
40:47
SPEAKER_01: And the idea there was like as microservices would take off, um, you would want to like uh understand the network calls API because your function calls were becoming API calls, right?
40:58
SPEAKER_01: So your previous generation solutions like new rally, cap dynamics, which were APM, were not very effective.
41:04
SPEAKER_01: So we built a network-based observability platform, uh, which was a hard problem, like going from packets to really understanding the application semantics.
41:12
SPEAKER_01: Um, that stream processor took a while to build, and as soon as it was built, I quit.
41:17
SPEAKER_01: We actually we gave Penn a stake in the startup.
41:20
SPEAKER_01: Your brother was yeah, so he was a how did he cut?
41:23
SPEAKER_01: So he was in Pennet well.
41:24
SPEAKER_01: So both of us came to Penn.
41:26
SPEAKER_01: Uh he came in.
41:27
SPEAKER_01: You know, he came in like a couple of years after me.
41:30
SPEAKER_01: Um I started actually Pen in 2009.
41:33
SPEAKER_01: 2008, I came to the US for my job first time.
41:36
SPEAKER_01: Like I it was I was still at Alcatel Ucent, and that's when I first realized that I have to be here.
41:41
SPEAKER_01: Right.
41:41
SPEAKER_01: 2009 I joined Penn.
41:43
SPEAKER_01: 2012 is when we, my brother, uh became a research assistant in the same group, and we started working on this idea.
41:49
SPEAKER_01: We coded for a couple of years, right?
41:53
SPEAKER_01: And 2014 is when we uh spun out.
41:56
SPEAKER_01: 2015 is when I came to the Bay Area, early 2015, and then we got funded.
42:01
SPEAKER_02: Understood.
42:02
SPEAKER_02: Amazing journey, right?
42:03
SPEAKER_02: A lot of resilience.
42:05
SPEAKER_01: I think you probably inherited from your parents or yeah, it was risk-taking, like in a way, like I was like, what's the worst gonna happen?
42:14
SPEAKER_01: I can always go back and get another get a job.
42:17
SPEAKER_01: I mean, that that was never like there was always good fallback options, right?
42:21
SPEAKER_01: But the idea was like, can we go on this non-linear growth trajectory, which is only possible if you're doing your own startup?
42:29
SPEAKER_01: Uh, and that's what we did.
42:30
SPEAKER_01: So there was a lot of resilience being a first-time founder with no like industry experience.
42:34
SPEAKER_02: No banking I I believe it would have taken you quite some time to raise funding.
42:40
SPEAKER_01: That's right.
42:40
SPEAKER_01: So we had raised from like friends and family, we had raised from some pen professors, we put in money, city of Philadelphia invested.
42:47
SPEAKER_01: We also took money from National Science Foundation, NSF for my first startup.
42:52
SPEAKER_01: So overall, we had like 450k to half a million uh out of the gate.
42:58
SPEAKER_01: Um then we incubated the product, tried to figure out the right market to build a company.
43:04
SPEAKER_01: Initially, we thought SDNs would be the place to do all this observability and application monitoring, but then we had to go into the microservices market slightly different.
43:15
SPEAKER_01: We came to the Bay Area in 2015, and right after that we found VCs who could understand what we were doing and see the trends with Docker microservices take you.
43:24
SPEAKER_01: It was very similar to the market we are seeing now with Code Rabbit.
43:27
SPEAKER_01: So there was a lot of chaos because your observability was broken as you went into this high velocity cloud environments.
43:35
SPEAKER_01: And that's where we built a product, the company, it went from zero to like a million ARR in the first quarter of selling.
43:43
SPEAKER_01: So it was very quick once the product hit the market.
43:46
SPEAKER_01: It was like, and we were getting so many meetings, strong inbound over even over there.
43:50
SPEAKER_01: But then what happened was um a couple of things.
43:54
SPEAKER_01: One was like we got really good inbound offers for acquisition for a first time founder, and it could be life changing.
43:59
SPEAKER_01: Yeah, life changing many.
44:00
SPEAKER_01: Right.
44:00
SPEAKER_01: So we had multiple companies show up.
44:03
SPEAKER_01: Then we went with Nutanix, which was public at the time, but then it had three other options.
44:07
SPEAKER_01: We also had like an option to continue building the company with Series A.
44:11
SPEAKER_01: But the cap table wasn't very optimized.
44:13
SPEAKER_01: Like there was a lot of dead stock.
44:15
SPEAKER_01: Pen had stock, which wasn't helping us.
44:17
SPEAKER_01: Like even founders had given up a lot in the first round of funding.
44:20
SPEAKER_01: So the best time option there was to just exit at that time.
44:24
SPEAKER_02: Yeah.
44:24
SPEAKER_02: And obviously you are in the the right place, right?
44:28
SPEAKER_02: You can build again.
44:29
SPEAKER_02: That's right.
44:30
SPEAKER_01: Then we continued now, second time around, then the challenge.
44:34
SPEAKER_01: So this was back in 2021 during COVID, which was also a very interesting time to start a company.
44:41
SPEAKER_01: Like the India was still shut down to find talent.
44:46
SPEAKER_01: SF was like a dead town.
44:47
SPEAKER_01: Like my first startup was in SF, but second one we did in Walnut Creek, where I used to live in the East Bay, uh in a co-working space.
44:54
SPEAKER_01: So the talent was very hard to find because everything was so badly shut down and everyone had to just be remote.
45:01
SPEAKER_01: So we found talent in Eastern Europe at that time.
45:06
SPEAKER_01: And that startup was right before Gen AI happened.
45:10
SPEAKER_01: And that time people still had appetite for uh complex infrastructure projects.
45:15
SPEAKER_01: So we had a theory that beyond observability, people will need controllability.
45:19
SPEAKER_01: Like, how do you go and prevent all these cascading failures using control systems and observability?
45:26
SPEAKER_01: So we built a platform, it was called Aperture.
45:28
SPEAKER_01: The idea was like how do you do smart load shedding and load management to prevent that your applications do not crash during a Black Friday spike or something?
45:37
SPEAKER_01: And that platform works really well, but then the market just shifted completely towards AI.
45:46
SPEAKER_02: No, I think what I learned from your journey is like be at it and you don't know when timing will work for you.
45:55
SPEAKER_01: Yeah, right.
45:55
SPEAKER_01: A lot of things are not under your control.
45:56
SPEAKER_01: Like you can always build cool time.
45:58
SPEAKER_01: I mean, things which are under your control, you can do really well.
46:00
SPEAKER_01: I mean, of course, I'm a builder, so I can build products, so that's not the hard problem.
46:04
SPEAKER_01: The hard problem I've always found is how do you market things and how do you sell them, the go to market, the product market fit exercise.
46:10
SPEAKER_01: I mean, that's also like a lot of it is like luck, yeah, being at the right place at the right time.
46:16
SPEAKER_01: Like these inflection points happen rarely, they happen once in a decade.
46:20
SPEAKER_01: But if you're building a startup and trying to create your own wave, that's what my second startup was.
46:26
SPEAKER_01: It's very hard.
46:27
SPEAKER_01: It's much easier to ride a wave.
46:29
SPEAKER_01: The first one we were riding a wave of Kubernetes, microservices taking off, and observability got pulled in.
46:35
SPEAKER_01: Over here, we are riding a coding agent wave.
46:38
SPEAKER_01: In the second one, there was no wave.
46:39
SPEAKER_01: We were trying to like show the people like what the best way of doing something was, but there was a lot of education involved.
46:45
SPEAKER_01: So, but but as an entrepreneur, you can never time the wave.
46:49
SPEAKER_01: That's right.
46:50
SPEAKER_02: To time it.
46:51
SPEAKER_01: It's hard, and some of these waves are like false start.
46:53
SPEAKER_01: Look at web three and all.
46:54
SPEAKER_01: I mean, not going anywhere.
46:56
SPEAKER_01: Like some of these things seem like like things that will be massive, but they don't end up that way.
47:01
SPEAKER_02: Because by the time, for obvious reasons, you you know there's a wave, you're already too late to the wave.
47:08
SPEAKER_02: That's that's my belief.
47:09
SPEAKER_02: So you can only know the early signs of a wave if you're riding it.
47:13
SPEAKER_02: If you are sitting on the shore, once it climbs at the top, then only you that's where I would feel you have to be in the arena.
47:21
SPEAKER_01: You really have to be on the ground.
47:23
SPEAKER_01: Like if if I wasn't doing in the second startup, even though it was not a great success, um, I would not have figured out this use case.
47:33
SPEAKER_01: I mean, it was really finding the use case in the arena.
47:35
SPEAKER_01: I mean, same thing happened with Slack, if you remember.
47:37
SPEAKER_01: They started out as a gaming company, but then the internal project took off.
47:41
SPEAKER_01: Um, so if you're really not on the ground, like I don't think you even will discover these opportunities.
47:46
SPEAKER_01: But once you find something, then it's about pivoting.
47:48
SPEAKER_01: Like there's always a sunk cost fallacy of like sticking around with old ideas and tarpet ideas.
47:53
SPEAKER_01: Sometimes you just have to be brave enough to even kill the bad ideas and prioritize um the things where you will create non-linear value quickly.
48:01
SPEAKER_02: Yeah, even I would say it's a very brave signal because Cold Rabbit was like an independent project and at a one million year, but so many projects are at a million revenue.
48:12
SPEAKER_02: So shutting down your own company and merging it with Cold Rabbit, because as an entrepreneur, you always love your own ideas.
48:19
SPEAKER_01: Yeah, that's that's always a tricky part I have found.
48:22
SPEAKER_01: Like you always fall in love with these ideas and you want to keep them alive for longer, but then the market decides that, like, you don't get to decide on what wins.
48:30
SPEAKER_01: I mean, we just present the ideas to the market.
48:32
SPEAKER_01: If it's not sticking, then it's as I said, it's some cost fallacy.
48:35
SPEAKER_01: And a lot of founders make the mistake, they will be like sticking around with a bad idea for many, many, many years.
48:40
SPEAKER_01: Try a slow death.
48:41
SPEAKER_01: Like, it's always good to like fail fast in these kinds of things.
48:45
SPEAKER_01: If it's not gonna work, uh, it's not gonna work, right?
48:48
SPEAKER_01: I mean, or you have to you basically have to pivot.
48:50
SPEAKER_01: Yes.
48:50
SPEAKER_01: And pivoting is very normal.
48:52
SPEAKER_01: Even my first startup, we as I said, started with SDN, dabbled with security angle for a while, none of those worked, and then when microservices were taking off, we were able to put a narrative behind it.
49:05
SPEAKER_02: What were some of the advantages when you were a founder second and third time?
49:09
SPEAKER_02: What things became easy and what still remained difficult?
49:12
SPEAKER_01: First of all, it's a very different time uh than 10 years back.
49:16
SPEAKER_01: Like a lot of the things, the way you run the business, I mean everything gets rewritten.
49:20
SPEAKER_01: Like the way you were doing go to market in the first was very content.
49:24
SPEAKER_01: Now, content is very different now.
49:26
SPEAKER_01: Like people don't go to medium.com and read blog posts.
49:28
SPEAKER_01: I mean, it it videos do way better.
49:31
SPEAKER_01: Like, I mean, everything changes, like so.
49:33
SPEAKER_01: So you cannot just take exact same ideas from your startups and translate them here, but you can like still do first of principles thinking on distribution and product.
49:43
SPEAKER_01: For example, like for this startup is all PLG.
49:46
SPEAKER_01: And by the way, this one I would credit more gur on this.
49:49
SPEAKER_01: I mean, I would not have really built it as a PLG business if it was me because I've not done it, but now I'm realizing that he did the first principle thinking, like he treated this market as a prosumer market.
50:00
SPEAKER_01: But my first startup was very different.
50:02
SPEAKER_01: So, yeah, some of the ideas don't translate well, some do.
50:05
SPEAKER_01: Um, and new mistakes get made.
50:07
SPEAKER_01: Sometimes you're overconfident uh about things, and and um sometimes um uh and that second time founder overconfidence always like backfires.
50:18
SPEAKER_02: So so one is the the coding market, which is the which leads to a code review market, both are growing very rapidly.
50:26
SPEAKER_02: But the second market is also vibe coding, which is lovable, emergent, and so many the criplet.
50:32
SPEAKER_02: So, where do you think that market is heading?
50:36
SPEAKER_01: So the the vibe coders are still going to be successful in a way, like a lot of people are now empowered, they don't have to go through the normal um Git process.
50:46
SPEAKER_01: Um they can just ship to production because those kind of things you um you can just vibe it to production, right?
50:53
SPEAKER_01: Uh but in any serious setting where you have real customers, uh revenue, uh SOC to compliance, whatever, the code reviews are going to stay.
51:00
SPEAKER_01: So both markets will exist.
51:02
SPEAKER_01: I think wipe coding is here to stay as well.
51:04
SPEAKER_01: Um, especially with like internal apps, those kind of things you can build quickly, websites.
51:08
SPEAKER_01: Right.
51:09
SPEAKER_01: You think the wipe coding market will still continue to right.
51:12
SPEAKER_01: The question is like, would it all become like part of the same system?
51:15
SPEAKER_01: Like can Clark code also subsume a lot of the lovable?
51:18
SPEAKER_01: Um, that is where it is like, or are the best in class solutions there continue to thrive.
51:25
SPEAKER_02: So so would Code Rabbit be also a solution to people building on lovable and emergent and other companies like Replet?
51:33
SPEAKER_01: We don't see that right now.
51:34
SPEAKER_01: We don't even have any customers who are let's say lovable customers using Code Rabbit.
51:39
SPEAKER_01: Like most of our customers are cloud code, cursor users in the past, uh GitHub Copilot, and so on.
51:46
SPEAKER_02: So so uh serious development hasn't happened, I assume, on that part, because if CS code was written on it, there would have been review.
51:56
SPEAKER_01: That's right.
51:56
SPEAKER_01: So most of the companies which are like, I mean, any company that is like sending shipping to real users will do a code review.
52:02
SPEAKER_01: Okay.
52:03
SPEAKER_01: I mean, that's where we are.
52:04
SPEAKER_01: And that other market is just a parallel market, of course.
52:07
SPEAKER_01: That's a massive market as well, and it's gonna continue growing.
52:09
SPEAKER_01: But we are not, we don't see ourselves playing in that market for any time soon.
52:13
SPEAKER_02: You you mentioned in the beginning of the podcast, very interesting, that every problem will become a coding problem.
52:20
SPEAKER_02: That's why the time is traded.
52:22
SPEAKER_02: Can you uh spend some more time on it?
52:25
SPEAKER_01: That's what we have been seeing.
52:26
SPEAKER_01: Like, if you look at even what uh OpenAI is now seeing the same thing, like Codex, you might have seen their ads on all the airports.
52:33
SPEAKER_01: I mean, it's becoming a consumer tool now.
52:35
SPEAKER_01: Um, and it's been integrated into Chat GPT, it's the same surface.
52:38
SPEAKER_01: I mean, and there's a good reason for that.
52:40
SPEAKER_01: Like the models are only capable for the uh depending on the surface area you provide to them.
52:47
SPEAKER_01: Like in the past, it had been MCP2s and all, they'd become more capable with more tools.
52:52
SPEAKER_01: Um, but now where we are seeing is like code generation is actually the best skill for the for the agents.
52:57
SPEAKER_01: They don't need anything if they can just generate code.
52:59
SPEAKER_01: Yeah, like they can run it, write CLI commands or actual programs to do work.
53:04
SPEAKER_01: Um, and that that's where the agents have been the most effective.
53:06
SPEAKER_01: And by the way, we knew this in back in 2023.
53:09
SPEAKER_01: So, what we are now seeing happen with the coding agents is what Code Rabbit was doing.
53:13
SPEAKER_01: And then we uh realized that potential.
53:15
SPEAKER_01: Um, and that is clearly what's happening in the market.
53:18
SPEAKER_01: Like you are seeing all the um not just the software development, these agents will not be just limited to software development, but you would be doing your slide decks through code, your Figma designs through code, your um um um entire operations, like for example, you are running get uh Google Ads through code changes now, it's all GitOps.
53:37
SPEAKER_02: Yeah.
53:38
SPEAKER_02: So so what are the problems that still will not be solved by code?
53:44
SPEAKER_01: I mean, and that's a good question in ways.
53:47
SPEAKER_01: Like I think it all boils down like the the you because you these these surfaces and code you're giving the agents really an agency to execute, right?
53:55
SPEAKER_01: Either whether it's information retrieval or um even like validating their work, it's much easier to do an eval loop when it's code, right?
54:02
SPEAKER_01: Um then and and and they can hill climb there, like that's where exactly what is happening.
54:07
SPEAKER_01: Anything you can translate to code and you have the clear test case against it, um, even the design you can test against everything.
54:13
SPEAKER_01: So you can hill climb.
54:15
SPEAKER_01: Anything you cannot hill climb is or translate a coding problem where the human taste comes in.
54:19
SPEAKER_01: That's where you need to be um, you can use the agents to accelerate, but then humans are still needed.
54:25
SPEAKER_01: Those are the places where the data is not there, there's no RL loop.
54:29
SPEAKER_01: Right.
54:29
SPEAKER_02: So, what are the other big problems that you think entrepreneurs could should focus on?
54:34
SPEAKER_02: Where is the market headed?
54:36
SPEAKER_01: I think it's a most amazing time.
54:38
SPEAKER_01: Of course, there's a lot of anxiety with like, I don't know, cloud code or codex taking it all, but I would say um people are just like very much limited by imagination on the actual product experience the public is gonna consume.
54:51
SPEAKER_01: And these are just infrastructure PCs, is what I see.
54:54
SPEAKER_01: This is just Kubernetes at this point.
54:55
SPEAKER_01: Like, what do you choose to build on it is is where the value is gonna accrue.
54:59
SPEAKER_01: The value is gonna be up the stack.
55:01
SPEAKER_01: Like it's still about the UX, the DX, like which products people um bring it part of their daily life, right?
55:10
SPEAKER_01: And some of these surfaces are still like wide open.
55:13
SPEAKER_01: Like, look at Slack.
55:14
SPEAKER_01: There's no good AI in Slack right now, right?
55:16
SPEAKER_01: And that's very logically the next place where everyone's spending time, collaboration is happening, but there's no great AI over there.
55:23
SPEAKER_01: So there's a lot of opportunity in like building new abstractions around as a company as code.
55:28
SPEAKER_01: So we have built all this tooling internally, but we see that someone from Code Rabbit could go and start a company around um your marketing as code or something, right?
55:36
SPEAKER_01: Um, so a lot of great ideas right now in the space.
55:40
SPEAKER_02: Yeah, what what do you think about the infrastructure and inference space?
55:46
SPEAKER_01: I think those companies are doing really well, especially with the open models.
55:50
SPEAKER_01: Um, like we ourselves use Base 10, for instance, for some of the open models we've started trying.
55:56
SPEAKER_01: Um yeah, I mean, there's an infinite, I would say, appetite for more computer at this point.
56:03
SPEAKER_01: Um, especially with open source models getting better.
56:06
SPEAKER_01: So these players, if they can figure out an efficient way of serving that inference, all the way from caching and batching and whatever, right?
56:13
SPEAKER_01: I mean, there's a lot of scope and potential there.
56:16
SPEAKER_01: Of course, the cloud companies are coming from behind, but not moving that fast.
56:20
SPEAKER_01: So either like it's a model lab that are building the super optimized inference tracks or these smaller players right now.
56:26
SPEAKER_02: So so do you think there is scope for specialized inference?
56:30
SPEAKER_01: Like somebody building inference for specialized for voice or there are many ways to specialize, even like things like latency, like some use cases, background agents are not latency sensitive.
56:41
SPEAKER_01: So you could really have an inference that can be slow but cheap, but high quality.
56:47
SPEAKER_01: I mean, for example, like even our use case somehow is not very latency sensitive, right?
56:52
SPEAKER_01: So so there's a lot of optimizations possible there.
56:54
SPEAKER_01: Then there's like, yes, of course, more real-time.
56:56
SPEAKER_01: So it's all about like that dose trade-off, like quality to performance and all that.
57:02
SPEAKER_01: So yeah, a lot of scope.
57:03
SPEAKER_01: Even there's a lot of scope for slow but cheap AI, but high quality.
57:09
SPEAKER_02: And uh for Code Rabbit, right?
57:12
SPEAKER_02: What does the next 12 months look like in terms of product, in terms of new market?
57:17
SPEAKER_01: Right.
57:17
SPEAKER_01: So we're like now becoming a platform, what we're calling as like agentic change management, right?
57:22
SPEAKER_01: So just like your Git has been source soft source code management, right?
57:26
SPEAKER_01: Um, and the market is becoming a change management market.
57:30
SPEAKER_01: Like the agents are the ones doing the changes.
57:32
SPEAKER_01: Like humans can start with a prompt, or these agents are not even getting prompted.
57:36
SPEAKER_01: More and more agents will be background, most of your tokens in 12 months are going to be background agents.
57:41
SPEAKER_01: Agents autonomously waking up every day and deciding whether they want to like take a customer feature or fix a data dog alert or even write a blog post on your change log.
57:52
SPEAKER_01: They'll open a pull request, right?
57:55
SPEAKER_01: And then that's what you have to manage.
57:58
SPEAKER_01: So one is like the quality of that pull request before even a human comes in, has to meet the bar.
58:03
SPEAKER_01: Secondly, humans need some sort of a help to triage and prioritize the work.
58:08
SPEAKER_01: So your linear Jira are getting disrupted.
58:11
SPEAKER_01: All that planning is now happening in pull requests.
58:14
SPEAKER_01: So you're not filing tickets, you're filing pull requests now.
58:17
SPEAKER_01: Your ideas are code, right?
58:19
SPEAKER_01: And now at that point, you first of all figure out like whether I need to ship a customer feature or whether I need to uh ship a um security issue uh fix or something, right?
58:30
SPEAKER_01: Then the third thing which can happen in 12 months is explainability, which I think is even a bigger opportunity than observability ever was given.
58:38
SPEAKER_01: There's a lot more people who need explainability than what observability the market was.
58:43
SPEAKER_01: Um because every problem, as I said, is becoming a coding problem.
58:47
SPEAKER_01: And the the biggest difference between like which agents are successful are the ones which can convince a human that their work can be trusted right now.
58:56
SPEAKER_01: So a lot of people can open slot PRs, but most of them get closed because you can't trust those outputs, or it's gonna take a lot of effort to validate their work.
59:05
SPEAKER_01: So this is where another AI has to come in.
59:07
SPEAKER_01: It has to really show the user on how the work is grounded and what all agents looked at, like whether it's through their thinking traces.
59:15
SPEAKER_01: So there is a way to like instrument these agents if you can control end-to-end, or like what we do in code review, we reverse engineer from the work we are able to understand how what the semantic layering of that work looks like so that you can build trust quickly.
59:29
SPEAKER_01: Like so, those are the things in the 12 months.
59:31
SPEAKER_01: I think this whole agent change management is going to emerge as a new category of products where you have to manage the changes done by the agents and humans provide the judgment where needed, which is given.
59:42
SPEAKER_02: Do you think um it's it's a winner-take all market?
59:45
SPEAKER_02: Or there are multiple.
59:47
SPEAKER_01: Where we are, I believe to a large extent, yes.
59:50
SPEAKER_01: Because if you look at the um coding agents, they are going to be fragmented and commoditized.
59:56
SPEAKER_01: Like as I said, that open source models are picking up, people have their cults, pref preferences around different harnesses, surfaces.
1:00:05
SPEAKER_01: So in every company you go to, you will have four to five coding agents people are using, just like how you had editors in the past.
1:00:13
SPEAKER_01: And then you have background agents.
1:00:14
SPEAKER_01: So that market is not going to be winner takes all.
1:00:17
SPEAKER_01: It's very clear it's gonna be highly fragmented.
1:00:20
SPEAKER_01: Right?
1:00:20
SPEAKER_01: Whereas when it comes to code review, given it's a central install-based team-based dynamic, you're only going to have one guardrail solution where you're gonna program everything on what you care about.
1:00:30
SPEAKER_01: Their checklist will be defined there and enforced over there.
1:00:35
SPEAKER_02: Would you also enter into then policy that uh or you are already there into a policy segment, like what are the uh, you know, uh the company policy, the GRC market?
1:00:48
SPEAKER_01: It's a natural extension too.
1:00:50
SPEAKER_01: And not at this time.
1:00:51
SPEAKER_01: We started doing some of the work around security and compliance now, so that's the first thing.
1:00:56
SPEAKER_01: Uh, new add-on product we are launching in a couple of days from now, um, which does continuous posture management as bomb and so on.
1:01:03
SPEAKER_01: But some of this other policy stuff we are like not for like there's a lot more other opportunities, like it's a very interesting point.
1:01:10
SPEAKER_01: Like, I mean, right now it's not just a land grab happening in go to market, but even what to build, there's so many options adjacent to us which can we can start chipping away.
1:01:19
SPEAKER_01: Yeah, thank you for that.
1:01:21
SPEAKER_02: I love the conversation.
1:01:22
SPEAKER_02: Congratulations on that you around and all the lecture.
1:01:25
SPEAKER_02: And uh looking forward to doing a part of something in the future.
1:01:29
SPEAKER_02: Thanks for being here and having me here.