Latest / The Index Podcast / Context.dev: How Angry Emails Built a 400-Customer AI Startup
Transcript
- 0:07SPEAKER_02: One quick word from our sponsor, Triton One.
- 0:10SPEAKER_02: If you're building on Solana, your app is only as fast and as reliable as the infrastructure underneath it.
- 0:17SPEAKER_02: At Triton, we built the open source read layer for Solana ourselves.
- 0:22SPEAKER_02: The streaming engine that pushes data the moment it hits the chain, the transaction pipeline that lands your trades when everyone else is timing out, all of it running on bare metal in over 20 data centers worldwide, served by the team that wrote it.
- 0:34SPEAKER_02: We power some of the biggest names on Solana, but with pay as you go, anyone can start.
- 0:39SPEAKER_02: Sign up, get an endpoint, pay only for what you use, head to Triton.1, serious infrastructure for serious builders.
- 0:45SPEAKER_02: I've worked with Triton for years at this point.
- 0:48SPEAKER_02: They've been around since the very beginning of Solana.
- 0:50SPEAKER_02: They've contributed open source software.
- 0:52SPEAKER_02: That's the core of the infrastructure of Solana that makes everything work since day one.
- 0:56SPEAKER_02: I couldn't recommend working with a better team.
- 0:58SPEAKER_02: Triton.
- 1:10SPEAKER_02: Is that how I say your first name?
- 1:12SPEAKER_00: Uh Yahya.
- 1:14SPEAKER_02: Yahya.
- 1:15unknown: Yeah.
- 1:15SPEAKER_02: Yahya Bakur.
- 1:17SPEAKER_02: Did I get that?
- 1:19SPEAKER_02: Yahya Bakur, founder of context.dev, a web context API that turns any website into clean, structured, real-time data for AI agents and software products.
- 1:32SPEAKER_02: Backed by Y Common Air.
- 1:34SPEAKER_02: Yahya is building the infrastructure that lets AI agents and developer teams pull accurate, fresh web data from brand assets to full site crawls through a single API.
- 1:46SPEAKER_02: And they're already powering products like Minsklify and Daily.dev as well as Klarna.
- 1:52SPEAKER_02: Super cool to have you on the show.
- 1:54SPEAKER_02: I really appreciate your time.
- 1:55SPEAKER_02: Thanks for being here.
- 1:56SPEAKER_02: Thank you for having me.
- 1:57SPEAKER_02: So before we get into more about what context.dev is, tell me your story.
- 2:02SPEAKER_02: Like what is the reason you get out of bed in the morning to do this?
- 2:06SPEAKER_01: Honestly, it's a ton of fun and it's quite fulfilling.
- 2:08SPEAKER_01: Like I think I've done quite a I don't know.
- 2:11SPEAKER_01: I've done quite a bit so far in my life where I did like different uh companies, bootstrapped companies, took a bunch of jobs, and this is by far the most fulfilling thing I've ever worked on.
- 2:22SPEAKER_01: So it's just it's very fun to build a data product if you if you're that type of person.
- 2:26SPEAKER_01: And then it's very fun to see what people build on top of your product.
- 2:29SPEAKER_01: So like we have 400 customers right now.
- 2:32SPEAKER_01: They constantly post about their building, they mention how they're using us, whether we're powering like a small feature or a core part of the product.
- 2:39SPEAKER_01: I think it's just fun.
- 2:41SPEAKER_02: 400 customers.
- 2:42SPEAKER_02: That's that's huge.
- 2:42SPEAKER_01: How long have you guys been in business?
- 2:44SPEAKER_01: A year and a half, but we doubled last month.
- 2:46SPEAKER_01: So that is uh that's amazing.
- 2:48SPEAKER_01: That's amazing.
- 2:50SPEAKER_02: That's incredible.
- 2:51SPEAKER_02: Because I remember when when you know Joe introduced us, you weren't there yet.
- 2:57SPEAKER_02: Number one, you didn't have 400 customers, and and two, you just literally like hired your first team members.
- 3:04SPEAKER_02: Um because like you've been basically built up to that first.
- 3:08SPEAKER_02: I'm I'm I'm venturing to guess the first 200 by yourself almost.
- 3:12SPEAKER_01: Yeah, the first 200 were an absolute grind.
- 3:14SPEAKER_01: I think literally the first customer took a month, and then the first 10 took like three months.
- 3:19SPEAKER_01: But with most product-led companies, they you tend to see this curve where you know it just takes a while to build some sort of critical mass, and then once you hit it, you just grow like grow like crazy because you kind of become a default option, or at least part of the conversation, or part of every conversation.
- 3:35SPEAKER_01: So I don't even think we're there yet.
- 3:36SPEAKER_01: I think there's still like quite a bit more to go.
- 3:38SPEAKER_01: Like Resend is a very good example of like I think they grew in, you know, it took them two years to get to 9,000 subscribers, and then one year afterwards to get to like 90,000, which was insane.
- 3:48SPEAKER_01: Wow.
- 3:49SPEAKER_02: Yeah.
- 3:49SPEAKER_02: Yeah.
- 3:50SPEAKER_02: Just huge, huge growth.
- 3:51SPEAKER_02: I I'm curious, what has this discovery process felt like for you as a founder?
- 3:57SPEAKER_02: Because you went from one customer to 10, right?
- 4:00SPEAKER_02: And that that zero to one to ten part, uh, I find is always like the hardest battle to find the signs of product market fit.
- 4:07SPEAKER_02: Like, am I even getting this right?
- 4:08SPEAKER_02: You know, are we getting any are we even directionally going in the right direction with what we're trying to build?
- 4:13SPEAKER_02: How how has that looked for you and felt?
- 4:15SPEAKER_02: And is it all like data or some of an intuition?
- 4:18SPEAKER_01: Like, walk me through what that is for you.
- 4:20SPEAKER_01: I mean, at the beginning, there's no data, it's like pure noise.
- 4:23SPEAKER_01: And you know, I was actually working on multiple projects at the time, so it's just one of them.
- 4:28SPEAKER_01: The problem is all of them kind of got stuck in like the 1k per month range, which sounds great.
- 4:33SPEAKER_01: Like, there is something there, but you need to pick what to double down on because you can't just work on a bunch of things at once.
- 4:39SPEAKER_01: So this was the only project where I'd have like these massive companies signing up and then getting like pretty pissed at how bad the data quality is.
- 4:48SPEAKER_01: They reach out to me and they'd send me like these angry emails, and I was like, wait, like, why did this person care so much or care enough to write this long email?
- 4:55SPEAKER_01: That means there's probably something here.
- 4:56SPEAKER_01: So I guess that was the signal that was like, okay, I should probably focus on this.
- 5:01SPEAKER_01: I killed the rest of the projects and then just focused on this one.
- 5:03SPEAKER_02: Oh, that's interesting.
- 5:04SPEAKER_02: So the first signal is just people were pissed it didn't work.
- 5:06SPEAKER_02: They like got they got bought in on the the value prop and then it didn't deliver and they were mad.
- 5:13SPEAKER_02: How did you handle that?
- 5:15SPEAKER_02: Like, how did you fix it?
- 5:17SPEAKER_01: I mean, then you just make one person happy at a time because people would ask for different things.
- 5:21SPEAKER_01: So I just picked one customer or one user at the time who was just the most pissed.
- 5:25SPEAKER_01: And it's like, okay, so if I add these things, you'll be happy with it.
- 5:29SPEAKER_01: Like, yes.
- 5:30SPEAKER_01: And then I actually added them and they ended up paying.
- 5:32SPEAKER_01: And you just do you kind of have to force the first 10 like that.
- 5:35SPEAKER_02: And did you did you keep it, was it from a barrier to entry standpoint, was like free to sign up?
- 5:40SPEAKER_02: There wasn't really like any any uh paywall in the beginning, or how did you think through that in the beginning?
- 5:45SPEAKER_01: I've experimented with a lot.
- 5:46SPEAKER_01: Like we tried paywalling at the beginning and it does work.
- 5:49SPEAKER_01: The problem is that then you have much higher churn because you're kind of misleading people because they don't really know what they're about to pay for, right?
- 5:57SPEAKER_01: So like you might have some insane churn because someone who thought they were getting like a marketing tool to generate assets doesn't realize it's actually an API.
- 6:05SPEAKER_01: So that was like a problem.
- 6:06SPEAKER_01: Um, I found that just doing a free free trial or like free credits at the beginning.
- 6:13SPEAKER_01: So like a certain amount of credits, when they run out, you just subscribe and then get uh constant renewals.
- 6:17SPEAKER_01: That works best.
- 6:18SPEAKER_01: Because then people know exactly what they're buying, churn is super low, and they'll give you feedback on why they didn't buy because they actually tried it out.
- 6:25SPEAKER_01: So you kind of get that feedback loop, which you wouldn't if you pay well at the beginning.
- 6:28SPEAKER_02: So, what is the thesis behind context at dev?
- 6:31SPEAKER_02: Like just taking a step back because we kind of jumped right back into how you got from zero to where you are right now.
- 6:35SPEAKER_02: I mainly because I just was blown away that you have 400 customers, and it's been like a month since we first talked.
- 6:40SPEAKER_02: That's nuts.
- 6:40SPEAKER_02: So congratulations.
- 6:42SPEAKER_02: Thank you.
- 6:43SPEAKER_01: 200 a month ago, so it's not it's only a double.
- 6:45SPEAKER_01: It's not like uh Yeah, but still, that's a lot.
- 6:48SPEAKER_02: That is a big inflection point.
- 6:49SPEAKER_02: I mean, when you think about product market fit and what that looks like on a chart, it looks like that.
- 6:54SPEAKER_02: You know, like you're you're starting to get there.
- 6:57SPEAKER_02: But yeah, like what is the problem you're solving?
- 6:59SPEAKER_02: What's the thesis behind what you're building?
- 7:01SPEAKER_01: So the problem, okay, everyone's building agentic products now, and they're building agents, and agents are out on the in the world.
- 7:08SPEAKER_01: They're going to need data to do anything useful.
- 7:10SPEAKER_01: It's the same thing with products in general, even pre-AI.
- 7:13SPEAKER_01: A product that was super useful is one that gave you normally had to connect to the internet in some form and pull data.
- 7:19SPEAKER_01: What we're building is the infrastructure for agentic products or agents in general.
- 7:24SPEAKER_01: So if you need any type of data from the internet, you would use us first.
- 7:28SPEAKER_01: Let's say you need to set up like monitors on the internet, you would use us for that.
- 7:32SPEAKER_01: Or you want to do like some massive crawl, you would use us for that.
- 7:35SPEAKER_01: So that's kind of the direction we're heading, is we're infrastructure for web data.
- 7:39SPEAKER_02: Yeah, well, let's dig into that a little bit because you know when I use like cloud code or something, you know, it can access the internet and find stuff, but why does it need why should I use context?
- 7:49SPEAKER_02: Is it not just like my cloud code session is some other use case?
- 7:52SPEAKER_02: What's the primary use?
- 7:53SPEAKER_01: So your cloud code session is you using it locally, right?
- 7:56SPEAKER_01: So just use your local browser to scrape like whatever info you want.
- 8:00SPEAKER_01: The problem is it's gonna take time and you don't have any scale.
- 8:04SPEAKER_01: It's kind of the same thesis of why sandboxes exist.
- 8:07SPEAKER_01: You know, I have a machine, why do I need the ability to spin up a thousand sandboxes?
- 8:11SPEAKER_01: It's so you can get that sort of scale.
- 8:13SPEAKER_01: So if you wanted to scrape 100,000 URLs to aggregate data, structure it, and then compare it against each other.
- 8:19SPEAKER_01: If you use your local machine, it's gonna take, I don't know, 12 hours, maybe a day, maybe two days, depending on how it's gonna navigate those websites, extract the data, et cetera.
- 8:28SPEAKER_01: With us, it takes like minutes, which is kind of the value prop.
- 8:32SPEAKER_01: And then if you actually want to deploy it into the world and charge people for it and build products around it, then you definitely need a service like that.
- 8:39SPEAKER_01: Because you can't just have cloud code running on your box to handle you know 10,000 concurrent customers.
- 8:45SPEAKER_02: Give me an example of how some of these customers are using it.
- 8:49SPEAKER_02: If you can give me like a case study, that'd be kind of helpful.
- 8:51SPEAKER_01: Yeah, yeah, for sure.
- 8:52SPEAKER_01: So the most common use case at the very beginning of the company was like personalizing onboarding for people.
- 8:57SPEAKER_01: So it's like the moment you log in, the more information that these products or companies have about you uh upon your sign up because you have a corporate email, the higher your conversion rate for onboarding is.
- 9:08SPEAKER_01: And then what the more magical the product just looks and feels.
- 9:11SPEAKER_01: So that was like an easy use case.
- 9:12SPEAKER_01: Afterwards, we started populating.
- 9:15SPEAKER_01: A lot of people use us to populate context on customers over time.
- 9:19SPEAKER_01: So let's say you maintain a knowledge base on your customer, so your agents actually have access to what a customer does, but then they update their website, they add a new product, they add a new service, they update their pricing.
- 9:29SPEAKER_01: Um, you don't want that to have the customer remember to go in and notify you that that happened.
- 9:34SPEAKER_01: So you can just detect it because it's on the internet and uh pull that data in.
- 9:38SPEAKER_01: Let's say you wanted to go, uh you wanted to build an agent that could really easily pull the pricing tiers off any website.
- 9:46SPEAKER_01: There's websites in hundreds of different languages, they put their pricing on different pages.
- 9:50SPEAKER_01: The data might little be fragmented across pages.
- 9:52SPEAKER_01: So, how do you do that efficiently?
- 9:54SPEAKER_01: You can give us a schema and we would pull it, we would structure the data for you.
- 9:58SPEAKER_01: So you effectively turn any website into an API.
- 10:01SPEAKER_01: Um, these are all use cases that people actually use us for right now.
- 10:04SPEAKER_02: That's super interesting.
- 10:05SPEAKER_02: I'm thinking of a side project I've been working on.
- 10:07SPEAKER_02: I I recorded some content about it just yesterday.
- 10:10SPEAKER_02: I've been playing around with ChatGPT's voice control for Codex, which has been awesome.
- 10:15SPEAKER_02: Like I can just walk around my room and talk to it and it just does stuff.
- 10:18SPEAKER_02: You know, and I I loaded up the desktop app with a bunch of skills that we built at our company to make like an agent team run efficiently and get good results.
- 10:27SPEAKER_02: And then I just set it off on this project, which was building an agencorkflow for sourcing and analyzing and underwriting multifamily real estate investments, like a whole pipeline because I'm building out my portfolio.
- 10:41SPEAKER_02: And I used it as like an excuse to demo how voice works, and I did like a video on it, it'll come out soon, you'll see it.
- 10:47SPEAKER_02: But what it what what I'm thinking about is I, you know, I've got these APIs that I'm paying for that pipe the data in about like a specific area.
- 10:57SPEAKER_02: And I think there's a bunch of publicly available data that I could probably use context for to get better enrichment on those properties.
- 11:05SPEAKER_02: And so basically I would give context the structure of the schema of how I want it formatted for my workflow, and then it would and then I would just ping the API and it would go get it from whatever sites I pointed at.
- 11:17SPEAKER_01: Exactly.
- 11:17SPEAKER_01: And what hinds of all the heavy lifting of like figuring out how to navigate, what pages to hit, how to pull the data.
- 11:23SPEAKER_01: Um it's very interesting product to build because whenever we it's it's hard to talk about use cases because you're building infrastructure.
- 11:31SPEAKER_01: So it's the same thing as like what do you use sandboxes for, what do you use EC2 for?
- 11:35SPEAKER_01: It kind of can be used across a lot of industries.
- 11:37SPEAKER_01: So that's exactly what we're seeing with our customers.
- 11:40SPEAKER_01: It's like very random use cases that are very high value in these companies where they're just happy that they have infrared to do it.
- 11:48SPEAKER_02: Well, it's interesting to me because it's very much like Google, right?
- 11:53SPEAKER_02: Like you're indexing sites, data on sites.
- 11:56SPEAKER_02: And you know, Google obviously had like a search algorithm that ranks the results to get you like the most relevant result when you do a query.
- 12:02SPEAKER_02: But context.dev is like that, but for the agentic age and for giving agents context to do jobs for us.
- 12:09SPEAKER_02: I mean, am I making the right connection here or am I telling it off?
- 12:11SPEAKER_01: But that's kind of what it feels like.
- 12:13SPEAKER_01: The idea is eventually we would build our own index.
- 12:15SPEAKER_01: Right now, it's just really good infrastructure on demand that kind of scales up.
- 12:19SPEAKER_01: We do it all on the spot to get you the freshest data, but yes, eventually we will need to build our own index.
- 12:23SPEAKER_01: That's kind of the next version of this.
- 12:26SPEAKER_02: Yeah, so you'll have to index the whole internet, basically.
- 12:28SPEAKER_01: Yeah, I mean, you don't need the whole internet.
- 12:30SPEAKER_01: Um, I think a lot there there's already a few players in the space.
- 12:32SPEAKER_01: I think there's like seven total.
- 12:34SPEAKER_01: And you definitely don't need the whole internet, you just need a some small subset.
- 12:38SPEAKER_01: And uh the internet in terms of just text is a lot less data than you'd expect.
- 12:43SPEAKER_01: It's measured in hundreds of terabytes rather than like some extreme numbers.
- 12:46SPEAKER_01: So it's not like unfathomable.
- 12:48SPEAKER_02: And it really is just text.
- 12:49SPEAKER_02: That's it.
- 12:50SPEAKER_02: It's just text-based.
- 12:51SPEAKER_02: It's very interesting too, though, because I I think you know, I was just having a conversation with a friend about this this morning.
- 12:57SPEAKER_02: There was this big lawsuit that came out that got settled with between Anthropic and all these authors.
- 13:02SPEAKER_00: Okay.
- 13:03SPEAKER_02: And what it came out was that Anthropic downloaded a bunch of literary works from a pirated website uh that like was all these without paying for it, right?
- 13:14SPEAKER_02: So they and then they trained their model, the the Claude on it.
- 13:19SPEAKER_02: And lawyers like filed a tort you know litigation and won.
- 13:24SPEAKER_02: And they settled for like$1.5 billion.
- 13:26SPEAKER_02: Each author only got like three grand.
- 13:28SPEAKER_02: Like the guy who spearheaded it, the author that spearheaded it, literally like spent three years on this and only got like three.
- 13:36SPEAKER_02: But he felt good about the fact that he kind of stood up for their them them.
- 13:40SPEAKER_02: But what the court ruled was the the part of that that caused the settlement to have to happen, the thing, the rule that they broke was pirating it.
- 13:48SPEAKER_02: If they had just paid, like who knows how much it would have been, but it would probably have been less than$1.5 billion if they just paid for the books and scanned them and then trained the model, that was totally fair use because of copyright law.
- 14:01SPEAKER_02: You know, I think there's a big conversation around scraping the web and using it for training data and who does it belong to and, you know, and how does that work?
- 14:07SPEAKER_02: But it it would seem kind of obvious to me that if your system is like superior at getting that data, not only can you feed it to agents so they have their latest context, but you can also use it for training models.
- 14:17SPEAKER_01: You definitely can.
- 14:18SPEAKER_01: Right now we're the I mean the main use case we're focused on is like infra for products because it kind of narrows down your customer base a little bit because you can go extremely wide with this without even uh much effort at all, right?
- 14:29SPEAKER_01: Like everybody needs data from the internet in some way, shape, or form.
- 14:32SPEAKER_01: On the whole like uh scraping litigation stuff, uh yes, that is like a big concern whenever you're building an index of the web, is like the authority or the type of the website that you're pulling the information from matters a lot.
- 14:44SPEAKER_01: And then if it's like, let's say you're you accidentally scrape like a social network that is not not like one of the famous ones, like a lesser-known one.
- 14:53SPEAKER_01: So you don't realize it's a social network.
- 14:55SPEAKER_01: And then somebody posts a comment with like a book that is you know not supposed to be there.
- 15:01SPEAKER_01: Um, you have to be very careful about how to pull that.
- 15:03SPEAKER_01: So that's why we're kind of doing that as a next step rather than doing it uh from the start.
- 15:07SPEAKER_01: But it is something I've thought about a lot.
- 15:09SPEAKER_02: What's something that has surprised you in the last, say, like 90 days building this, like that sort of caught you by surprise anything?
- 15:17SPEAKER_01: Yeah, I think a lot of typical advice around building companies is usually you know, focus on one thing, go all in, uh, make those people really happy, it's maybe expand one step at a time.
- 15:29SPEAKER_01: Like a lot of traditional advice has just been wrong, I guess is the best way to put it.
- 15:33SPEAKER_01: Like whenever we ship something new, the customer base just seems to absorb it faster than I can ship it.
- 15:37SPEAKER_01: Like quite literally, I will we don't even have that many customers, right?
- 15:41SPEAKER_01: And it's an API product, so they're not supposed to be on dashboard every day.
- 15:45SPEAKER_01: But like we ship this new feature, and then before I announced it, like within 20 minutes, somebody had already created like we ship monitors and somebody had already created like 500 almost immediately.
- 15:54SPEAKER_01: And it's a very interesting experience when you ship something new that you're worried like whether it meshes with the product or not, and then suddenly you're just your customer base just absorbs it right away.
- 16:04SPEAKER_01: Um, that was unexpected.
- 16:06SPEAKER_02: How many people on the team?
- 16:07SPEAKER_02: And are you guys using agents to build the product?
- 16:10SPEAKER_01: Uh four people, yes, we're using agents very heavily, especially on like research loops and eval loops to see okay, how can we improve performance?
- 16:17SPEAKER_01: Because what now that we have we do about 30 million requests a month right now.
- 16:21SPEAKER_01: Wow.
- 16:22SPEAKER_01: So you can always take a sliver of prod traffic and rerun it locally to see, or rerun it with agents, like a sandbox, to see can you improve performance in some way?
- 16:30SPEAKER_01: Like what happened during this period.
- 16:32SPEAKER_01: Um, this is stuff that would have taken engineers like weeks back in the day.
- 16:35SPEAKER_01: Like one engineer would have taken weeks to do performance improvement.
- 16:38SPEAKER_01: Now we can do it in like an hour, which is great.
- 16:40SPEAKER_01: So you can have agents looking at metrics, they see whenever, okay, at this time of day for this customer, there was like a spike in latency or the spike in CPU usage that would have toppled over the box or something else.
- 16:52SPEAKER_01: You take that sliver of traffic, rerun it locally as long as there wasn't ZDR or anything on it, and actually improve the performance and then ship a fix out like that.
- 17:01SPEAKER_01: It's great.
- 17:02SPEAKER_01: I think one thing that's very worrisome is because we have so much volume, we just can't go down.
- 17:07SPEAKER_02: Yeah, you know, you guys are never sleeping.
- 17:10SPEAKER_02: Yeah, yeah.
- 17:11SPEAKER_01: So you you just pager duty.
- 17:14SPEAKER_01: Yeah, you need to like have a very good test set up.
- 17:16SPEAKER_01: You need to do a lot more testing than we used to, especially automated testing.
- 17:22SPEAKER_01: But yeah, pros and cons.
- 17:23SPEAKER_02: So you're hiring for DevOps.
- 17:25SPEAKER_01: Our DevOps is our agents.
- 17:27SPEAKER_01: But no, we have four people now, uh including me, and then just split roles.
- 17:32SPEAKER_01: But yes, DevOps is.
- 17:33SPEAKER_02: But that's pretty cool though, that you're using your agents basically as your DevOps monitoring system and improvement, right?
- 17:39SPEAKER_02: And I'm curious to learn.
- 17:40SPEAKER_02: So these loops, like what does that actually mean?
- 17:42SPEAKER_02: What is that you've got an agent loop running?
- 17:45SPEAKER_02: What does that mean?
- 17:47SPEAKER_01: So let's say um there's this wonderful tool called Axiom, Axiom.co.
- 17:52SPEAKER_01: I do not know how they make money because it's it's insane.
- 17:54SPEAKER_01: You can ingest like a thousand gigabytes of logs per month for 25 bucks a month.
- 17:59SPEAKER_01: It's insane.
- 17:59SPEAKER_01: I recommend everybody watching this just plug in them um instead of New Relic or Datadog.
- 18:04SPEAKER_01: But uh, we send all of our logs there, and then we set up these custom monitors via the MCP.
- 18:10SPEAKER_01: Uh we just say, hey, you know, like imagine how you used to set up monitoring on your product.
- 18:14SPEAKER_01: You'd go, you'd find like a line of code, you'd be like, I want to monitor this, I want to monitor that.
- 18:18SPEAKER_01: Now you just tell your agent to set it up and create monitors on the web, and you go ahead and keep track of like CPU usage or RAM or latency or like data quality if suddenly you have a spike of what looks like bad data.
- 18:31SPEAKER_01: And then you just take whatever traffic happened during that time, all the requests, all the responses, you load them in locally, and then you just tell your agent, hey, figure out how to improve this.
- 18:40SPEAKER_01: Or you give it some type of specific goal.
- 18:42SPEAKER_01: It it really depends on what went wrong during that time or what degraded during that time.
- 18:47SPEAKER_01: But you give it a goal and you just let it run for a while.
- 18:49SPEAKER_01: And I mean, you do have to steer it sometimes because it'll go on like weird paths that may not be useful, but uh that normally means your setup is lacking context in some way, shape, or form.
- 19:01SPEAKER_02: That's really interesting.
- 19:02SPEAKER_02: That feels like a whole business in and of itself to me when I hear you say it.
- 19:05SPEAKER_02: I'm like, other people will want this.
- 19:07SPEAKER_02: You know, yeah.
- 19:08SPEAKER_01: It's uh I mean there are businesses I I I know of a few that are trying to do this, the AI SRE, you know.
- 19:13SPEAKER_01: So that that's basically the loop they're doing.
- 19:17SPEAKER_02: Very cool.
- 19:17SPEAKER_02: But you guys just hacked it together yourselves using Cloud Code or something like that.
- 19:23SPEAKER_02: Yeah, yeah, pretty much.
- 19:25SPEAKER_02: So I mean, as a lean as a team of four people, like what are the agentic systems other than that?
- 19:29SPEAKER_02: Are there any other things you're doing to kind of get hit the scale and like support all those 400 customers?
- 19:34SPEAKER_02: You know, you that's a lot.
- 19:35SPEAKER_01: So emails I handle, we handle customer support ourselves.
- 19:39SPEAKER_01: Like we have an AI chatbot on the website.
- 19:41SPEAKER_01: It's primarily just there because we get like people asking questions in other languages now, and it's just too slow to respond.
- 19:47SPEAKER_01: Especially if they're asking, like, do you have an API for this?
- 19:50SPEAKER_01: You know, if I were to do it myself, I'd have to go copy the message, translate it, come back, translate it back.
- 19:56SPEAKER_01: Um, so we have like these AI support agents, but we usually step in if it's not resolved within uh three question turns.
- 20:04SPEAKER_01: And then on the customer side, I think that's about it.
- 20:08SPEAKER_01: We have agents running for different things, but we try to only automate something after it's it's proven to work like manually.
- 20:15SPEAKER_02: Given where you're at, you just hit this inflection point.
- 20:18SPEAKER_02: What's next?
- 20:19SPEAKER_02: Like what are you what are you prioritizing and how are you coming up with those priorities?
- 20:24SPEAKER_01: So there's a lot we can do, which is both a blessing and a curse, right?
- 20:27SPEAKER_01: There's like infinite directions we can take this because it's still very young.
- 20:31SPEAKER_01: And we've kind of proven that it if we ship not as I mean, there's a point where you've just shipped too much stuff, but everything we're shipping is kind of correlated, it's all infra, we're packaging it in a really good way, and customers are absorbing it and really enjoying it, right?
- 20:45SPEAKER_01: They're getting a lot of value out of it.
- 20:47SPEAKER_01: Um, so once you as a business establish those facts that your customers will enjoy you shipping more stuff to a certain degree, um, and you can go in different directions, and also customers will be happy with that, you can go in even more directions.
- 21:01SPEAKER_01: So for us, right now it's just about making sure we're doing the right next thing, which we're about to add more company data, we're about to add like certain amounts of people data, and you can also blow up the scope there.
- 21:13SPEAKER_01: So trying to keep it focused on hey, you can use this as infrastructure for your products.
- 21:18SPEAKER_02: And I'm curious, you know, you're in you're in Y Combinator.
- 21:21SPEAKER_02: Are you still in?
- 21:22SPEAKER_02: Is the is that cohort still happening?
- 21:24SPEAKER_02: And what's that been like?
- 21:26SPEAKER_01: It is it is definitely still happening, and it's been an experience, man.
- 21:30SPEAKER_01: It it's a lot of fun, like and and very tiring at the same time, right?
- 21:35SPEAKER_01: Because you're just working all the time.
- 21:37SPEAKER_01: You're setting these outlandish and and just crazy goals every two weeks.
- 21:42SPEAKER_01: And the the crazier part is you actually get close, right?
- 21:46SPEAKER_01: Like a lot of people don't hit the goals, but the fact that you even set such an outlandish goal and you're kind of uh going full sprint towards it means you get way further than you would have without it.
- 21:55SPEAKER_02: Um it's fun.
- 21:56SPEAKER_01: Yeah, it's going well.
- 21:57SPEAKER_02: Are you seeing any other like AI companies or agentic companies that you're excited about in the cohort or or elsewhere?
- 22:02SPEAKER_02: Like what are you getting any alpha?
- 22:05SPEAKER_01: I don't think I can say, but yes, there are some people that are absolutely killing it, and they're also extremely cool people too.
- 22:11SPEAKER_01: Um, because we kind of started the batch together, all of us, and we're like, this is gonna be fun.
- 22:16SPEAKER_01: And then you go through it.
- 22:16SPEAKER_01: It is fun, it's just super tiring.
- 22:18SPEAKER_01: It's much uh much more tiring than you expected, right?
- 22:21SPEAKER_01: Um, for some people, you know, their entire business is sales led, so they're just traveling uh every two days across the country.
- 22:27SPEAKER_01: Uh for me, my business is product led, so I don't need to do that.
- 22:30SPEAKER_01: I still need to take calls all all the time, but I don't have to actually do as much traveling.
- 22:35SPEAKER_02: Well, that's good.
- 22:36SPEAKER_02: Um, and as far as you know, what it took you to get here, what are you know, I think you you were at Amazon before this in Sunrun.
- 22:43SPEAKER_02: What are some of the lessons you learned there that have kind of helped inform your strategy today and like help you get to the this far with this company?
- 22:50SPEAKER_01: Amazon was great to see what proper engineering looks like.
- 22:53SPEAKER_01: Like, you know, a lot of people have different gripes with Amazon's engineering culture, but it it is legit, right?
- 22:59SPEAKER_01: Like you do learn how to be a good engineer there, you do learn how to do things at sort of gigantic scale.
- 23:04SPEAKER_01: Um, you need to forget some of that when you're building a startup because it's just not relevant, but it does help with building a product that's already doing you know 30 million requests a month and that's growing much faster.
- 23:14SPEAKER_01: So it helps to make sure we don't go down on the engineering front.
- 23:18SPEAKER_01: Uh Sunroom was just a lot of fun.
- 23:20SPEAKER_01: Like it was a solar company, but we were this AI team inside.
- 23:22SPEAKER_01: We're building products nonstop that would drive like a ton of revenue for the company.
- 23:27SPEAKER_01: And um it was in a field that sort of benefited from data, but also it I don't know, it was just a very interesting time there.
- 23:36SPEAKER_01: Um it really helped me figure out like how big companies work uh when you're kind of near budget this, but I guess I guess budget decision makers is a good way to put it.
- 23:47SPEAKER_01: Um so it helped with that side of things.
- 23:49SPEAKER_01: But yeah, I guess this is next.
- 23:51SPEAKER_01: And then I did stock alarm at some point, which was uh my first startup.
- 23:54SPEAKER_01: So that was fun.
- 23:56SPEAKER_02: Very cool.
- 23:56SPEAKER_02: So we're getting towards the top of the show.
- 23:58SPEAKER_02: I always ask this at the end, but what have I not asked you that I should have asked or that you wanted to talk about?
- 24:03SPEAKER_00: I think you covered it all.
- 24:04SPEAKER_00: I don't know.
- 24:05SPEAKER_00: I not nothing comes to mind.
- 24:06SPEAKER_02: Awesome.
- 24:07SPEAKER_02: Well, look, I really appreciate you coming on the show.
- 24:09SPEAKER_02: Super exciting to hear all the progress you've made, especially since the last time we we talked, and congrats on the big jump and 200 customers.
- 24:17SPEAKER_02: And I hope that I hope that growth continues for you, man.
- 24:19SPEAKER_02: That's really exciting.
- 24:20SPEAKER_01: Thank you.
- 24:20SPEAKER_01: Yeah, let's uh let's see what next month looks like.