He quit his job, went all-in on AI agents—then grew to 100K users & a $30M Series A in a year. | Soham Ganatra, Founder of Composio

Soham Ganatra (00:00:00):
Don't believe too much in what customers are talking about. If you're building a new product and if you have very strong intuition on why it will work. I would say this is one of those times where multiple products like, iPhone could come out. Which is if you go ask people, they won't really tell you like your product is what they need. But if you just go build it and show it, they would really get it. So my hunch with AI space is that you can't validate problems by asking people, because they won't get it. A friendly customer won't tell you to, like this doesn't make sense. I'm not going to use anything even close to this, right? They'll be like, okay yeah, you know, what? Maybe if you do this, you can go use it. Maybe if you can do this, maybe we can use it but you end up spending six months, right? And it's they're trying to do their best to accommodate you, in some way. But truthfully you don't want people to accommodate you. Because that's the whole point. 

Previous Guests (00:00:56):
That's product market fit. Product market fit. Product market fit. I called it the product market fit question. Product market fit. Product market fit. Product market fit. Product market fit. I mean, the name of the show is product market fit.

Pablo Srugo (00:01:08):
Do you think the product market fit show, has product market fit? Because if you do, then there's something you just have to do. You have to take out your phone. You have to leave the show five stars. It lets us reach more founders and it lets us get better guests, thank you. Soham, welcome to the show, man.

Soham Ganatra (00:01:24):
I appreciate you having me.

Pablo Srugo (00:01:25):
No, dude, I'm excited to have you. I mean, you're one of them. I mean, we're doing this new series. We've done, you know, Cluely, Wispr Flow, Hedra. Kind of these new, really hot AI apps. Obviously, AI is kind of taking over the world and I find a lot of them at the helm are young founders like yourselves that are just killing it. Frankly, I mean, you just raised like a $30 million Series A from Lightspeed a couple of months ago and as far as I understand. You have, how many users using your platform?

Soham Ganatra (00:01:53):
Yeah, at this point. Like more than 100,000 developers.

Pablo Srugo (00:01:55):
100,000 devs, crazy and just in your two-year-old company. So yeah, man, let's dive into that. I mean, we'll get. I mean, maybe just a first question, right? In the simplest terms possible, non-technical, as much as you can. What does Composure actually do?

Soham Ganatra (00:02:12):
Think about it this way. If you are a developer or if you are even a web coder. You're looking to build an AI agent and you want that AI agent to sort of work with real world, right? So you want the agent to essentially talk to Salesforce, talk to Gmail, talk to Outlook, do a bunch of different tasks on it, right? What you will realize is you go through a bunch of small issues you face like authentication and authorization problems, you face integration issues, then you face reliability issues, and all of this combined take a lot of your sort of bandwidth. What we offer is a single platform that essentially solves all of these problems for you. So as someone who's building any agent that is looking to interact with the world. You can just use Compose, SDK, APIs, whatever, and go live with real-world use cases very quickly.

Pablo Srugo (00:03:02):
And is this different than. I mean, in the world of maybe more typical SaaS. You've got these companies like Merge that came along and there's a bunch of others that are just these integration platforms, right? So instead of building your own integrations, you might rely on them. Is that what you're doing, but for AI agents? Like in this AI world?

Soham Ganatra (00:03:17):
Yeah, that's a cool way to think about it, right? Integration world in general has a bunch of these different companies. So your Merge, you have sort of Zapier and Tray's. What you could think of Composio in a way is like we do the same thing. In a very specific manner, specifically for agents and the LLMs.

Pablo Srugo (00:03:37):
Got it. Well, let's kind of start at the beginning. I mean, especially since there's not a long timeline, right? To go over, but what's your background? Where does the original idea for Composio come from?

Soham Ganatra (00:03:49):
Composio started sometime in 2023. This was basically me in a product manager role, engineering manager role in my previous company. Learning the difficulties, like typical sort of team faces while doing a bunch of integrations.

Pablo Srugo (00:04:05):
Where were you? What company?

Soham Ganatra (00:04:07):
So I was in a company called Bureau. Bureau essentially does fraud and risk management. We started with a very simple idea. Which is like, for every company out there. They essentially go through this hyper scaling phase. Where a lot of users sign up and some of these users are not the best of breed. And what you want to do is, you want to detect those users very early on. So we created a platform to sort.

Pablo Srugo (00:04:31):
And what stage of a company was it?

Soham Ganatra (00:04:33):
So I joined in as the first person in the company and I basically went through series A. And the hyper scaling phase ended up sort of at some point, we were like 50 plus people in 2023. When I was sort of, now they're series B and above. So yeah, they've been just growing phenomenally.

Pablo Srugo (00:04:57):
And did you have other jobs before this? Or were you in school before?

Soham Ganatra (00:05:01):
So before this, I started a startup. I started a startup in consumer AI. So my journey has been. I ended up doing my schooling in 2017. I started a startup back then. The startup was into a consumer sort of AI space. We were basically building chatbots. Ended up going towards enterprises, selling to larger enterprises, some of the largest ones in Asia. Exited my company back in 2019. I wanted something more chiller, because startups are exhausting.

Pablo Srugo (00:05:35):
Good big exit or good small exit? Or what kind of situation was that?

Soham Ganatra (00:05:40):
It was a small exit, essentially.

Pablo Srugo (00:05:44):
But still a positive sort of end to the journey.

Soham Ganatra (00:05:47):
Yes, I think like pretty positive. Because it's interesting learnings and very intense sort of one and a half years ended up teaching me a lot. Ended up sort of deciding that we wanted something a little more calmer. Decided to join as the founding engineer. Didn't anticipate like that was anything but calm, right?

Pablo Srugo (00:06:10):
I could have helped you there, man. I could have told you that, yeah.

Soham Ganatra (00:06:13):
Yeah, so to go through the same journey, right? So went through the same thing again for like three years. I went through the, this time was a bit more into let's say successful and so we went through the scaling part very quickly. Very early on, ended up sort of growing the team from zero to fifteen people. That was amazing, learning a bunch and yeah. Going through some of this whole sales promotion, product promotion. Did a bunch of different things in the meanwhile, essentially, at the company.

Pablo Srugo (00:06:44):
And what do you see? I mean, like November 2022, ChatGPT goes out. Obviously, it's a crazy moment. Were you playing with GPT a lot before that? Or was that kind of when you started to kind of go into this world?

Soham Ganatra (00:06:54):
Yeah, I was one of the early users of GPT-2. The beta users out there. It was horrible, by the way. Yeah, it was good but you don't really think like, okay, this is something that could actually be practically useful. It's just very interesting, so you play with it. Yeah, I think the perspective changed in those six months, right? Because, I've been thinking about this problem for a very long time, right? Which is like software connecting to another software is so painful and as developers, you kind of hate doing the work. But you have to, because it's one of the most integral things in your product. And so the idea was like, how do you solve for it? And practically you just can't. Because every software has a very different interface. Every other software has a very different interface and so, how do you connect these two interfaces that just look very different, right? And a bunch of these companies came up, right? That you talked about, and everyone's trying to do the same. But if you go to consumers and you go to people who use them. And you ask them, they all kind of hate it. Because it's just like the problem occurs. It's just in the six months period, they're all good. But after that, it's just like, why to do so much custom work to actually make it working for me?

Pablo Srugo (00:08:14):
Is it because? I mean, my non-technical understanding of it is, there's different levels of integrations and so you might get in like an integration to say HubSpot through Merge.dev or whatever. But then you decide, oh, I actually wanted to do this other thing and it's not supported. And so now you somehow have to, like you say, customize it and you end up doing a lot of the work anyways. And so you're like, what am I, you know? What are we paying for this thing for?

Soham Ganatra (00:08:33):
Exactly, right? So at some point it's almost comical but your NPS is like 10, 10, 10, 10, 10, and then as the time increases your NPL is drastically, starts decreasing. Because everyone starts discovering these ten other things they want to do with those integrations and they just can't. And it just becomes a very painful thing, right? So that is a problem, right? And my thinking was that back then, it's like LLMs could solve for it if we just had that magical core generation. That just allowed me to connect any software to any software. Been playing around with like, GPT-2 that sort of such a GPT-Da Vinci. All of those things coming and at some point I thought like, I felt enough confidence that, okay, this thing is improving at a rate. Well, if I make a bug today and maybe like a year or two years down the line. It will just be good enough. Well, like the tech would actually get to that point. Where I could maybe solve for this problem, and that's how we sort of ended up starting a company.

Pablo Srugo (00:09:37):
What did you need? I mean, maybe just again. Walk me through, what did you need the AI for? Given what you, because that's I mean, now you're doing. Obviously it's for AI agents, but I don't think it was at the beginning. So at that point, when you're seeing this stuff with GPT. Where does the AI fit into your product at that stage?

Soham Ganatra (00:09:53):
Sure, so think about it this way. So I have a interface on the left side and I have an interface on the right side. I want to connect it together. That connection requires customized code. So the way you think about this problem is that, typically with any integration product, like there's a thirty percent or twenty percent product component. There's a sixty to seventy percent service component for near down the line. Because a lot of this customization is actually powering this product and finally everything is coming together. And the idea was this sort of service component or where some human is writing code is really crappy. Nobody really likes to do it and it makes the end-to-end experience very bad. What AI could potentially solve for was, the automation of that service component. So the idea in the beginning was, that what if we built a product and that product could essentially have in some way an adjustable layer that. This sort of AI could come in and adjust basis of the interface on the left side and on the interface on the right side. And suddenly through basically some magical thing. I could connect both the interfaces together without any services, without anything coming in between.

Pablo Srugo (00:11:05):
When you talk about the service. What do you mean by the service layer? Is that the maintenance piece or what's that piece?

Soham Ganatra (00:11:11):
It's a bunch of people coming in and adjusting the code to fit the interface. So I will tell you an example. Let's say I am looking to get like name, email, and maybe address of the user from my CRM, right? And Salesforce has some way to integrate, right? And so maybe they have this two fields, name and email. but they store address in maybe address one, address two, pin code and something like that. Maybe HubSpot does it differently where they have a single field address. Where the complete address is stored, and so when I'm trying to get this data into my database, right? Which maybe stores it, in something like address and pin code. Like those are the two fields. I had to do a bunch of transformations to essentially solve for it,  and those transformations are taught by a developer out there. Who's writing custom code, right? To essentially make sure that the data from Salesforce gets transformed in some way. Data from HubSpot gets transformed in some way, and what my thinking was. Was that if code generation could become really good. We could essentially have this magical thing that could understand, what the data is stored on the right side with. What is the data on the left side with, and it could essentially write these transformations, and everything on the fly. And if the transformations need to change, because maybe my database schema changed. It could be a very simple update that could essentially allow me to change those things.

Pablo Srugo (00:12:37):
And so once you see this. And you kind of see the potential for AI to do this in terms of integrations. How quickly do you leave the startup that you're at? How do you make that decision?

Soham Ganatra (00:12:47):
I think it was pretty much in a month or so. Because I think the way I operate is, if there's a problem. That I can't stop thinking about. At some point you understand that this is a problem that really matters to you and so I was just like, yeah, I really need to do this, right? I think I can't really work on anything else right now. That's all I'm thinking about is this problem. So that ended up happening very soon afterwards.

Pablo Srugo (00:13:13):
When was this, by the way?

Soham Ganatra (00:13:14):
This was somewhere around April and May, 2023. Because I saw just like coding becoming really good. I saw some of the weekend hacks that I ended up building, coming out to really good results and so that just gave me a lot more confidence. So the project kind of started as weekend hacks, right? It's just like, I used to hack on weekends. A bunch of these things with my friends and some of these things turned out to be pretty good. And so I was like, okay, this makes sense. I should maybe, you know, think about going full time with this, yeah.

Pablo Srugo (00:13:47):
And was it just you at the time? Or did you have co-founders right away?

Soham Ganatra (00:13:50):
No, so I had a co-founder who has founded a previous startup with. So it was like we both knew each other from a very long time and so it was like pretty easy. We both did a bunch of projects together and we both know each other since, last like twenty years now. So yeah, it was like with him essentially, and we both started it together.

Pablo Srugo (00:14:10):
And so what's your? Do you go out and once you decide to leave. Do you raise money right away? Do you start building for a bit? What's your kind of first move?

Soham Ganatra (00:14:18):
Yeah, so my co-founder was newly married and he just had a daughter. For him, it was imperative that we go out with some security. Because it's not something you want to do right after having a kid, right? We ended up going to VCs and we ended up raising some money. Right after leaving the job and starting up. Simply because we wanted some level of security in our mind. If this doesn't work out, we can still pay our salaries. At least some base level salaries.

Pablo Srugo (00:14:51):
How much did you raise?

Soham Ganatra (00:14:52):
We ended up raising close to $4 million, $4.5 million.

Pablo Srugo (00:14:56):
How was that? I mean, 2023 was an interesting time. Where the markets had actually kind of been crashing. Not a great time, but then AI was on the rise. How did you find? You know, raising for an AI seed stage company at that time?

Soham Ganatra (00:15:10):
I think it was not hard, not easy as you said. There was a mix of reactions. But also, we did have some reputation. Because of the previous things we had ended up building and the companies we were coming from. So I think those reputation, that reputational advantage kind of helped. Also I think 2023 was pretty wild in a way. Because it was, I would say months of extreme FOMO and months of on the VC's end, right? And months of extreme, oh, I don't want to invest in anything, right? 

Pablo Srugo (00:15:44):
Yeah.

Soham Ganatra (00:15:45): 
So I think we just got lucky with timing. I really would think, maybe if we would have raised couple of months later. We might not have been able to and like before might not have been able to. It's just like you need to raise at the right time. When you see things coming.

Pablo Srugo (00:15:58):
Were these VCs that you knew from before? Or did you just kind of run a process? How did you get around it going?

Soham Ganatra (00:16:03):
No, so I knew them from before. Some of them I knew from before. Some of them was introductions of friends and that ended up helping us out.

Pablo Srugo (00:16:12):
So now you have the money, you've got your co-founder, do you build a first product of this? Do you try and validate the problem? Or are you confident in the problem itself? What's the biggest risk at this stage in the company?

Soham Ganatra (00:16:26):
Yeah I, like, okay. So my hunch with AI space is that, you can't validate problems by asking people. Because they won't get it. So the advantage we had was, we kind of had seen the problem before. We were facing the problems in our previous ventures. So we understood it very quickly, and we kind of also understood that this is something definitely worth solving for. So problem validation was not top of the mind. I also don't think if I would have pitched it to twenty people. Let's say all of them would have gotten it, right? I would have probably been able to convince maybe four or five.

Pablo Srugo (00:17:07):
Why is that by the way? I'm just curious on it. Because the flip side is, it's funny with AI I do find that you're right in what you're saying of problem validation not being the problem I should say. Because it's like, a little bit of AI kind of feels like magic. So it's almost this if you build it, they will come, you know? If I could give you, you know, every single integration you ever wanted without having to do anything. Would you want that? Yeah, who wouldn't, right? But you're saying a lot of people might not have gotten it.

Soham Ganatra (00:17:32):
No, so exactly, right? So they would have just called bullshit on it essentially. Which is like, this seems too good to be true, right? And so that's the reason they would have not wanted it, right?

Pablo Srugo (00:17:43):
I see, okay.

Soham Ganatra (00:17:44):
Which is like, yeah, of course. Like what you're, exactly. Of course you give me free money, I'll take it. But like, could you?

Pablo Srugo (00:17:50):
So it was about. So really the key thing was making it work, building it.

Soham Ganatra (00:17:54):
Exactly, the key thing was, is the tag there? So we started building coding agents back in 2023. We were basically trying to figure out, how to build good coding agents back then. So we ended up testing out a bunch of frameworks. Agent was not even a thing back then, like agent became a thing I think a bit later. So for us, it was just like, we're trying to figure out tool calling. We're trying to figure out like, okay, how to connect this. I don't know, with GitHub and all. We're trying to almost make an AI engineer that could just do integrations. We're basically literally building like DevInfo integrations, right? And we're realizing that this is not a problem that is going to get solved, but that took a bit of time, right? I think that took like six months to. It took us six months to come to the realization that, this is not a problem technically that could get solved. Six months we had.

Pablo Srugo (00:18:42):
How many people were you at that time? In those six months?

Soham Ganatra (00:18:45):
We were like four people at this point.

Pablo Srugo (00:18:47):
Okay, so you kind of saved the money. You banked the money and you just, the four of you kind of huddled up trying to make this thing work.

Soham Ganatra (00:18:53):
Yeah, I don't think we even spent like $100K out of it.

Pablo Srugo (00:18:56):
Yeah, on four people, yeah.

Soham Ganatra (00:18:58):
So we had all the money. We just didn't want to spend it until we're sure, we're spending it on something, right? We also had customers. So we had like, again, I think good reputation comes in handy. Because you could just go talk to your friends and they're like, yo, if I could give you all the integrations, would you want it? And they're like, sure, yeah, of course and so that's how we started doing a POC with Glean. Glean was a pretty large company back then and they had integrations as a problem. So of course it made sense for them to do a POC. We ended up doing that. Yeah, so we are like a very large enterprise customer and looking like enterprise.

Pablo Srugo (00:19:32):
They're paying like a six-figure type POC material.

Soham Ganatra (00:19:35):
They were ready to pay more than that if it goes well, right? And what we realized while building it for them was, okay, this is not going to work. Clearly the tech is not there, right? I think, we are maybe two generations or three generations of LLM away. I still remember, I think we used to think GPT-5 would solve it. Maybe and like, we didn't know GPT-5 would take this long a time to come out. But, yeah, that was the hunch. I still think it's not a solvable problem today. Maybe like one generation away.

Pablo Srugo (00:20:06):
Oh, interesting.

Soham Ganatra (00:20:06):
Yeah, but like. Yeah, that's how we started, right?

Pablo Srugo (00:20:10):
So six months. I mean, you're six months in. It's probably what? End of 23', beginning 2024, you're realizing. Yeah, this thing's not. It's not gonna fly.

Soham Ganatra (00:20:17):
Yeah, this thing's not it. Yeah, exactly and at some point we decided. Okay, like, you know what? We have to figure out what the fuck we are going to do. So, one of the other problems we realized. While doing all of this was like, we ended up writing a lot of integrations for our agents and I was like at some point the FUBINI came on the team. It's like okay why are we building integrations for the old school software? Let's just build integrations for the new age software, right? And that's how we sort of got into the other like I would say the next gen of Composio. Which was like, okay let's just go build integrations for AI agents out there, right? And that was pretty cool, I would say. Because that was again a problem that we kind of realized we faced. We faced it extensively in fact, right? And it also made sense because, most people were yet to figure out this problem. Because we're playing with coding agents way back. I don't think many teams in the world were playing with that much coding agents back in those days.

Pablo Srugo (00:21:16):
And this problem, let me just understand this. This problem surfaces because you're trying to build coding agents that build integrations from, say, HubSpot to sales, HubSpot to X, or Salesforce to Y, or whatever. And in so doing, you need these agents to actually access those destinations as you're kind of running into this agent to whatever integration problem. You're probably one of the first people ever to run into it. Because who else? There's not that many people in the world building with AI agents and you know mid 2023 or so

Soham Ganatra (00:21:47):
Yeah and tool calling was not super stable. A bunch of other things were, not mostly done. And so like, apart from a lot of other problems we faced, while building those agents. This was one of the other, like this was, I would say one of the interesting problems we found and I was like, okay, let's just go solve this problem right now. Which is, how do I make it super easy for someone who's building. Maybe something like Composio? To essentially go live with their agents, right? And so that's how Gump came into the new generation of product, essentially.

Pablo Srugo (00:22:18):
And what's the difference? Fundamentally between building an integration from two legacy softwares, to each other and from an agent to access Gmail or to access HubSpot, or whatever it is?

Soham Ganatra (00:22:30):
I think a bunch of things change, right? So your old school integrations would be completely deterministic integrations, but when you think of hot agent tech integrations, they're more. So at that point in time tool calling was the only thing. So it was like okay how do you build like this JSON schemas that can take inputs? How do you make it super reliable? Because, LLMs can't understand complicated. Tools are complicated, like sort of integrations, essentially. How do you solve for authentication? Because, if I wanted to connect an agent to my Gmail, even today. If you do it without any tool out there. It will take you a long time because you have to just go solve the OAuth stuff first. You have to make sure that you need to, basically, you can do it like a bunch of hacky ways. But if you want to do it in production, like in a nice way. It's just so much hassle, right? And so we went through all of that. And at some point we're like we just need to use the same tech. But make that tech into a product rather than basically using it for our end product at some point.

Pablo Srugo (00:23:37):
But what I'm trying to get is. When you're talking about, you know, Gmail and I understand. I guess the deterministic versus non-deterministic piece, but in terms of the integration layer. Is it just that you have to now kind of take into account, when it's going to be called, how it's going to be called, when the agent's going to decide that it needs to use a particular tool. What are the pieces that you have to build that you maybe didn't have to build before? Or is it the difference between the two worlds of integrations?

Soham Ganatra (00:24:04):
Sure, so basically when you think about agentic integrations and how agents consume sort of like integrations. They consume it in a different way. Okay, so I'll give an example. So when I was thinking about, let's say no code integrations or integrations in the early age. I'm, let's say trying to create a lead on HubSpot. Maybe the lead requires a category and that category would be a dropdown. So that category would be shown to me in a UI. So I would say, hey, I can select a specific category of the customer. Maybe this is a sort of sales, maybe a sort of HR company or something like that, right? But when it comes to agents, you are not actually showing them a UI. You are giving them all the right APIs and the right APIs into this layer, that agent can understand. And for a long time, that layer was like a JSON schema. That you had to pass in sort of an LLM call, that the agent can consume and so the whole DX of how you do it changes drastically. And so it's not as much as a change in underlying core logic, but it's a really big change in the way a developer consumes the whole thing. So it's more of a change on the consumption side, but also because of now the change in developers' way of consuming how integrations are used with agents. There's a lot of changes on the way you think about an integration. So an agentic Gmail integration would be designed entirely differently than a normal sort of Gmail integration that is designed for humans to use. Because the way agent navigates Gmail is pretty different compared to the way sort of a human navigates Gmail. An example would be, okay, if I take an email out and that email has extremely long contacts. And if you feed it directly to the agent. The agent will just break, right? Because it just can't consume that much context, right? But a human could essentially keep consuming context, keep figuring out, keep going about it. So it's the way you design the API would be different. The way you think about an agentic sort of integration is entirely different and I think this kind of. There's a field of it, research field around it. This is agent computing interface, ACI, that you could look into, right? There are other ways you could think about it. Which is a lot of studies done around function calling benchmarks. Which is the technical way of looking at how sort of LLMs interact with different systems out there. So yeah, there's a bunch of work around it. But, yeah, it's like, how do I make it so that. When your iPhone is customized just for you, you use it so much better? How do I customize this, integration of Gmail just for an agent?

Pablo Srugo (00:26:50):
Right, no. That makes a lot of sense.

Soham Ganatra (00:26:51):
And how do I make it super optimal? Yeah.

Pablo Srugo (00:26:54):
And how does the team? As you decide to effectively pivot into this, is everybody aligned? All four of you are like, yeah, this is the thing? Or is there a little. Is there any misalignment or conflict internally about whether this is the actual thing you should do?

Soham Ganatra (00:27:08):
I don't think there was conflict, but like people didn't get it, right? I think it's, very hard to bet on some problem that maybe you saw, but like nobody in the world sees. If you Google search, oh, is anyone else facing this issue? Okay. No one else turns up.

Pablo Srugo (00:27:26):
No one's there, yeah.

Soham Ganatra (00:27:27):
Okay, yeah. This is to ahead of it's time, right? Are we actually betting on something real? And so there's always this conflict, but at some point. You kind of have to make calls like, okay, that's the bet I'm willing to die on, right? Which is, if this doesn't work out. It doesn't work out, right? And then, people don't trust the bet, but they trust you. And they're like, okay, I'll just ride with it. So yeah, that's how it went, essentially.

Pablo Srugo (00:27:54):
And then did you go into another kind of three or six months of just build phase? Or what was the next phase? What did the first half of 24' look like?

Soham Ganatra (00:28:01):
Yeah, so this time building was pretty easy. So it was like, oh, we know the problem. We have kind of solved it. We know how to go about it. So building took us like two months, right? Maybe, so it was like, I think we built. We got a product out by March or something. March end or something like that, right? Of 2024 and once the product was out. We started basically, this time we didn't want to do the mistake of like, getting a friendly customer. Because, a friendly customer is good and bad. It's good because it's easy to find. It's bad because, just end up giving you wrong signals maybe sometimes.

Pablo Srugo (00:28:42):
Were they too friendly? Is that what you found? They were too accommodating? Like, you know what I mean? So you didn't get that kind of real, like, dude, this is shit. I'm not going to use it. It was just kind of too much like, yeah, okay, we'll try it here, we'll try it there. Wanting to make it work.

Soham Ganatra (00:28:54):
Exactly, right? A friendly customer won't tell you to. They won't be like, okay, like, I'm not just going to use this. This doesn't make sense. I'm not going to use anything even close to this, right? They're like, okay yeah, you know what maybe if you do this you can go use it. Maybe if you can do this, maybe we can use it, right? And it's like yeah but you end up spending six months, right? And it's, they are trying to do their best to accommodate you in some way. But truthfully you don't want people to accommodate you. Because that's the whole point, right? Which is if you want real PMF you don't want people to use you. Because they know you or whatever. They should be just so sad about not getting anything out there. That as soon as they see you, they use you. It's very different.

Pablo Srugo (00:29:44):
It's so funny, man. You say this now and it's kind of clicking for me. Because I've seen this happen several times with repeat founders. Especially where, because they have these relationships, it is an edge. But sometimes a kind of weak value product goes way further than it should and they have these great logos, and they can say these big customers. What's happening is probably what you're describing. People that are like, yeah I know this guy, you know. CEO is like you just make it happen or CTO, or whatever. Someone really important just tell somebody else to go ahead and work with them. And ultimately, they just, all you're doing is taking longer to face reality. Because the market either is gonna love it or they're not

Soham Ganatra (00:30:23):
Yeah, I think like. So I used to do some trading, like in part time. But essentially, that's the thing, right? It's like market always tells you the truth at the end of the day. You just want to get to the truth faster, right? So, that was a learning, major learning for me. Which is, I understood. I never want to talk to a friendly customer ever. I just want to find people who don't know me and make sure that they like my product at the end of the day. And if they do, I think I'm just moving in the right direction at the end of the day. So that was the idea.

Pablo Srugo (00:31:01):
So how did you go out and do this with, you know, Q2 or whatever, 24'. You have the product. How did you go out and get new customers?

Soham Ganatra (00:31:10):
So there's this field of study around tool calling and accuracy around tool calling. So the thing is, if you want to integrate LLMs with anything out there, or agents with anything out there. You have to do what is tool calling. You have to use tool calling to connect an LLM to Gmail, for example and so it clearly was something every single user of ours or potential user of ours would go through. Which is like.

Pablo Srugo (00:31:36):
And this was, by the way. Sorry to interrupt, but this, just the background here. This was when GPT was putting out in ChatGPT, like these these tools. I don't know if they've gone anywhere, honestly, within ChatGPT. But there was this whole thing about a ChatGPT app and you call tools, and whatever. Was this a similar time frame?

Soham Ganatra (00:31:52):
It was similar, but it was actually before. Basically, what they inferred. It was, I think, quite a way before because, this is an abstraction they came up with. So that, LLMs could interact with systems outside, right? So LLM in itself is like a text in text out. But programming is like, the code is kind of structured and deterministic. So how do you go from text out to code, right? And so the layer between is, in a way. The way I think about it is to call it, right? So it allows LLM to generate a structured JSON that could represent here, what's the weather in New York today? And then that query goes to a function out there, an API out there that comes back with, hey, the weather today is this, and then LLM consumes that response. And basically shows you a text saying, hey, today the weather in New York is like very sunny, have a great day, right? Something like that. So it's a way LLM can interact with the system outside the sort of LLM itself and like basically get some data that it needs in real time. The way I think about any integration is, you want to interact with Gmail or HubSpot, or whatever. You have to go through tool calling essentially and that's the technical term. And what we realized was nobody was talking about it. Because tool calling kind of suck and, if you tried actually making something interaction back then. It won't work and so what we did was, we basically wrote a really nice blog and some code structure. And some examples out there on GitHub. We put it out and we said like, hey, here's some good ways you could improve your code. And it just hit the right audience. Because every person who was dealing with that problem was like, okay, this is so interesting. This is exactly what I wanted. Can I go and do some of these techniques to, maybe in my code base to improve my code? I like, they were very simple, very nice, described ways of improving tool calling and everyone kind of ended up liking it. It went viral on Reddit. It went viral at a bunch of different places and we got insane traffic because of it. And that's how we got our initial users.

Pablo Srugo (00:33:58):
When was this exactly?

Soham Ganatra (00:33:59):
I think this was somewhere around March or April last year, basically.

Pablo Srugo (00:34:02):
Oh, okay. So that was all the same timeframe.

Soham Ganatra (00:34:05):
Yeah, it was all the same timeframe. So basically this time I wanted to get independent users. So I started reading up on content and all of those things. I ended up doing some of these things in parallel to the product development, to make sure we start getting users out there. Yeah, so that went really well, I would say. Ended up getting a bunch of people to look into us, keep checking us out and try our product out. And that gave us a lot of feedback. I started just basically talking to a lot of them. In parallel also there's another way we went. Which was discord communities. So think about it this way, this is back, like a year back but discord communities for a lot of this agentic frameworks were kind of really big and you would probably have more than 10,000 people in some of these communities. Like developers who are building actual agents out there or who are  trying to build some interesting applications with LLM out there and basically I just went, and started spamming a bunch of them. And the spam was really nice. It's like, hey, here's an article we wrote, right? That also talks about tool calling, or here's an example, some really cool sort of demo we created, right? I like check it out and a lot of people really liked it, right? And so they started responding then, a lot of founders of those communities. I still remember CrewAI founder essentially put out a Twitter post saying, hey, this is one of the coolest things I've found today, and that was pretty good. That was the first time someone recognized us on Twitter and that got us a bunch of users. So I think things like those ended up helping out a lot. Because you kind of had to do a lot of work. You had to be compatible with CrewAI, for example. You had to do a lot of work to become compatible. You had to spend a lot of effort in creating that example. But then it hit the exact users that you wanted to hit and so that ended up getting us a lot of traction. And yeah, so that was how that came about.

Pablo Srugo (00:36:01):
One question I have is back then, where was the product? How many of these tool calling integrations did you have? How many tools were you able to kind of support?

Soham Ganatra (00:36:12):
We had very little, right? So we had maybe ten integrations or something like that, at that point, right? Because integrations are hard and if you want to sort of build integrations, it takes time. So we were trying to build integrations in parallel, but at that point, we have very few integrations.

Pablo Srugo (00:36:30):
And so what? The people that come in that early April, May 2024. What are they even using? I mean, they're stuck with just using the ten integrations that you support?

Soham Ganatra (00:36:39):
Yeah, exactly. But the way we went about it was, we essentially had this chat box on the website. That intercom's great that way. You can essentially pop up the chats at a bunch of places and so we used to have customized chat pop-ups on integration screens. Saying, oh, looking for more integration talk to me and I can help you out. So I ended up booking, I don't know ten to twenty calls a day for a couple of months and that was huge because I used to just.

Pablo Srugo (00:37:07):
Actually ten to twenty a day? 

Soham Ganatra (00:37:09):
Yeah, so some of them were no shows, some of them of course were like sure but likeye ah. But all in all I would say I averaged more than ten calls a day for a very long time, yeah.

Pablo Srugo (00:37:19):
And this was what, fifteen minute calls? People telling you what they wanted to see from the product?

Soham Ganatra (00:37:24):
Exactly, and it was very high. Okay, I would say it's very high signal, because you had a lot of people who just were trying AI for the hype around it or something like that. But you did meet a lot of very interesting people who told you about what they were trying to build and how potentially Composio could be useful or something like that, right? And so, yeah, that turned into a very good sort of product roadmap on insights for us. Which is how should we be thinking about the next few months? What should we be building? How should we be going about it? What are the things people want to see? Stuff like that, yeah.

Pablo Srugo (00:38:01):
And for the rest, the ones that where you had integrations that you supported. That they wanted to use. Were you charging or was it free to use? How did you set up the business model at the beginning?

Soham Ganatra (00:38:11):
No, so I think I never cared about pricing for a very long time. I think we still don't care that much about charging people per se. Because the way I think about it is, in the early stages of market. You just want to know where the pulse is at. You want to know what people find useful, what people want to see more, what are the things that are missing in your platform. I think for me, at least the way I used to think about it was a lot of these folks are independent developers who are doing personal projects. Because they're very interested in the new tech and to charge them right now is kind of a disservice to them as well as me. Because the value they get is, they want to build something cool. The value I get is, I get to learn from what they are building. To how to improve the product and pricing is something that we can all. Like getting value out of the product is something we can always do later.

Pablo Srugo (00:39:11):
How has that evolved? Do you just have a it's free to use up to this many API calls, and then it's some price per API call? Or what's the model?

Soham Ganatra (00:39:17):
Yeah, exactly. So we have a usage-based model. It's very generous, right? So for anyone out there, they can come start with Composio. They can actually scale to a pretty high volume out there without worrying about it. Without worrying about the pricing and then after that. We have something like, hey, here's a base tier, $29 plan that can get you this much usage and then if you scale beyond this use. Here's how much you pay for a million API calls or something like that.

Pablo Srugo (00:39:46):
And so then back to the storyline. You put out that blog, people come in, you post on Discord, whatever and then what keeps it going? Is it just pure word of mouth? Or do you have these other viral moments that kind of keep the slope going up?

Soham Ganatra (00:40:00):
Yeah we did programmatic SEO. So one of the things that I wanted to do was, I wanted to understand what integrations people want but like how would you know that? You have to talk to everyone but the other hacky way I understood I could know that was. I just ended up creating a lot of pages that just said, oh, integrate. I don't know, LangChain with HubSpot, right? And using Composio, and basically someone would come on that page, that would be a signal to me that. Okay, someone really wants to do HubSpot with LangChain and maybe I should make sure that I sort of practice that. So we did that, that worked really well until like. At some point, I think users are pissed, like, oh shit, you don't really have it. Why are you advertising that? And at some point we decided like, okay, we should not probably have random pages like this. So we removed it. But like in the early days, I think that helped a lot.

Pablo Srugo (00:40:58):
Give me a sense, of the numbers. How many developers, users did you get on a monthly basis in those first months? And then how quickly, like now you're over a hundred thousand. What did that year of growth look like?

Soham Ganatra (00:41:10):
In the early days, till like probably June, July. We didn't have more than I would say thousand users. It was super slow, because people who signed up and actually understood tool calling or building agents. And stuff were very, like people who kind of actually had use cases were far, and few. And a lot of folks did come on the website, and stuff. We had probably I think at some point in those days. Maybe like twenty thousand to thirty thousand sort of visits sometime around June, July, on the website. Which is pretty nice, because of a bunch of reasons like programmatic SEO and stuff. And so people who came on the website were many, but the people who signed up and actually used us were very few. I think because LLMs just didn't work and so this number didn't increase for a very long time. This number started increasing sometime in sort of the second half of the last year. When models just became better and you could actually build real world agents out there. That kind of worked and a lot of startups coming in, right? And so then it scaled very quickly to way more users, right? And way more people sort of building on top of us and so the number I think ideally scaled in like September. Sometime in September last year, started scaling then, right? Till then, it was just word of mouth. Okay, it's something good. I think it's useful, but I don't think my agents actually were. Do I really use this, right? Stuff like that.

Pablo Srugo (00:42:51):
So in September, you've got a few thousand users and now what is it? Eight, ten months later, you're at a hundred thousand.

Soham Ganatra (00:42:59):
Yeah, exactly. So then it started scaling pretty rapidly. Then I think the scale up to a point was, till I would say till MCP came up and MCP became big. Which is sometime around March or April this year. I think it was scaling at a scale then with MCPs, the scale became even bigger. Because MCP had a very good hype around it and so the other thing I realized was. You have to latch on to the good hype cycles in AI. Because they're one of the good ways you can find people. If the hype curve is right like if you're actually latching on to the right hype cycle and people actually find value from what you're building in that specific height cycle. It's a really cheap way to get a lot more users, right? And so we started getting way more users after MCP became big. And basically a lot more folks got a building on top of us. So those were the two transitions I would say in the last one year. Where we had a different scale of users coming in.

Pablo Srugo (00:43:58):
What is, you know? Because you're so in the middle of it. What is the big difference in your view between the MCP and kind of just the APIs you were dealing with before?

Soham Ganatra (00:44:10):
Yeah, sure. So essentially before MCP, it was like, okay, here are all the tools out there and you can go integrate those tools, right? With MCP, it just became for us. Here's an MCP compatible way, essentially the platform and the core integrations remain the same. The quality of those integrations remain the same, but you can use us via MCP also and so it just becomes a different way you could use this. And a lot of developers would end up using through MCP also. Would end up directly using Composio platform also. The difference though is now that we are going into sort of a higher abstraction level. The difference is now we're trying to sort of go a step above the integrations and we are saying, we'll also basically solve for reliability issues. Which is the biggest pain point in a lot of folks right now today. When getting agent live in production. Which is like, I want to connect my agent and I also want my agent to actually end up working after those integrations. And Composio is basically in a way, think about it. Collecting all of these skills. So think about it this way, when you drive a car and you drive a car today, and you drive a car months later. The skill of driving a car has changed. You have probably learned a bunch of different things. You drive a car better, but that doesn't happen with agents today. So agents don't learn. So if my agent is connecting to Salesforce and Salesforce has ten thousand different parameters. And it keeps messing up those parameters every time. It will keep doing that even like the third time, the fifth time, the thirtieth time, right? But with Composio, what we do is we essentially understand from your usage pattern, and we basically change the agentic tools itself. To make it more and more compatible for your use case. And so what we are doing is we are essentially trying to figure out, depending on the usage pattern. How do you best represent this tool? And it could be Salesforce, it could be anything else. So that in your use case, like, agent doesn't have to use as many tokens. Maybe it's far more sort of accurate. It's way better and this could be done on top of MCP. This could be done without MCP also.

Pablo Srugo (00:46:26):
Right, I guess so. I mean, that's what Composio does. But I guess I'm just curious on your take on MCP. Because you said, there's a lot of hype on it and for me as a non technical person. I mean, I read into it, but I'm curious on your take of, what's the big difference? What's the big thing you're supposed to get as a result of MCP being a thing? You know what I mean?

Soham Ganatra (00:46:44):
Oh, sure. So I can tell you my take on it, right? 

Pablo Srugo (00:46:46):
Yeah.

Soham Ganatra (00:46:47):
In the way I think about it. Okay, so when you think about integrations, there are always two kinds of integrations. Natively, you are actually thinking about. So let's take a product like ChatGPT, right? So the first integrations I would think about would be native integrations for ChatGPT. So it'd be like, hey, ChatGPT natively supports like this. I don't know, fifty applications. But then the second thing I think about is custom GPTs, right? It's like, oh, I can extend ChatGPT by basically this custom GPTs or by basically adding all of these other things to it, right? So one is integrations that basically the app developer is looking to support. Because they just feel it's native to the experience and then one is the kind of integrations that are, basically something that the user wants but it's not super native. But it's very specific to, hey, I want my ChatGPT to connect with Outlook. Because that is what I care about and so MCP enables the second kind of integrations. Which is, you could turn any sort of smart software into an extensible thing. So that as a user you could say, hey, I want my specific software like cursor to connect with Jira, right? Because Jira is what I care about or, maybe calendar to cursor. Because that's maybe what I care about, right? And so MCPs are great that way, because if every software becomes kind of an MCP client or maybe let's say good software that you are using. Maybe they become MCP client you have this advantage of connecting that client to a server. Which could be by Notion then you might say, hey, this software now connects to Notion, right? Because that's the integration I care about. So that's my take on it, which is it creates a lot of value by making every software into an extensible thing that users could decide what they want to connect it with.

Pablo Srugo (00:48:49):
And then maybe just to wrap it up here. Last quarter you raised, what I see as a $29 million Series A from Lightspeed. Tell me about how that happened?

Soham Ganatra (00:48:59):
Yeah, so I think at some point. Growing pretty rapidly, we wanted to sort of expand the team. We wanted to hire a bunch of research folks, right? And in AI, it's very expensive essentially. And so what you do need is you need more capital essentially to solve for it. I've realized this somewhere in the starting of this year. That, hey, we should be figuring out way to sort of raise some money. So that we can hire some folks that we really want to.

Pablo Srugo (00:49:28):
Cause you were how many people at this stage?

Soham Ganatra (00:49:30):
Well, somewhere around ten people. Ten to fifteen, I would say. Yeah, so ended up realizing that we need to sort of raise it somewhere around end of January, February. Because this is when most VCs are active. Because it's a start of new era and that they're looking to sort of get into new deals, right? Ended up reaching out to a bunch of folks, through our existing investors, through some of our friends, meeting a few of them. The race took somewhat close to two to three weeks and ended up starting to chat with a bunch of them in the beginning of that, sort of race cycle. Lightspeed was someone who we kind of knew from before. So my co-founder knows the sort of partner very well. I also know about them from different friends and so we had that personal relationship. So it became pretty easy, because when you trust and know the person. It's just like a way faster to sort of get into the details of it essentially and yeah, that's how we ended up sort of converting Lightspeed into digging in.

Pablo Srugo (00:50:40):
I mean, two to three weeks. Obviously very fast and it's still, $30 million dollars a day is a massive Series A. Do you have any any kind of tips or tricks, or tactics, or even just a high level thinking of how you should go about raising a Series A? Especially how you should go about raising one that is big and fast?

Soham Ganatra (00:50:59):
Yeah, many actually but okay. So I would say the first thing you do need is some level of hype, right? So what you need to figure out is right people should be talking about you and they should be saying good things about you. You should have customers that absolutely love you, right? And they should just, if any. I think at some point, a lot of our people we were trying to raise from ended up reaching out to a bunch of customers and they just heard some insane, amazing things. They were like, okay, this is the craziest customer calls we've been to. So, they just spoke so highly of the team and the product that it just like, it's a no brainer. I think that helps. Other thing is the story has to be right. So I think there are different kinds of raises. Like you could raise 15, you could raise 20. It's like you could raise 30. It's at the end of the day, valuation is not something that's objective. It's pretty subjective. So the story has to be something that you really think through and you are actually basically practicing it a bunch of times before going, and telling it to folks out there. So the way I did it was, I ended up practicing it with ChatGPT. Which is, I don't know. Something at least most of my friends now do. It's like I had all my transcripts with all the VCs in a single ChatGPT session and I used to say, what did I say wrong? Where should I have improved? What are the questions I was not prepared for? What are the questions I did? Maybe, good, but.

Pablo Srugo (00:52:35):
Give me the perfected now, the perfected, thirty second or minute story of. That you would have, you know, after all that, trial and error you now have.

Soham Ganatra (00:52:45):
It's been a while since I raised. I don't remember that story.

Pablo Srugo (00:52:47):
Not that long, right? Like what? Three months?

Soham Ganatra (00:52:50): 
The whole idea is this, right? Which is, the way I think about it is. You have essentially three layers, right? You have the foundation models, you have the in front between, you have the applications on top, right? And today what's happening is you have some other layers like agentic frameworks and stuff like that, right? And today what's happening is that the foundation models are coming to the top, right? They're trying to move further and further away to the top. What Composio does is, Composio essentially kind of enables foundation models and maybe at some point. Sort of becomes one of the de facto ways by which you can also take on the application layer. So think about it this way. The reason you interact with Salesforce and the reason you interact with any application out there. Is because, yeah, they have your data but There's nobody else who could break into that monopoly today. Because your sort of interaction is, very geared towards how you know Salesforce and how you think, basically you use Salesforce. How it has been fully customized for you. But what if Composio could essentially break that barrier and we could essentially become that de facto thing, where. Which, let's say we just know exactly your user pattern on Salesforce. We're able to completely customize it for you. We know exactly here are four fields in Salesforce that you care about and we are able to sort of give that to you at the right places at the right time, right? And so at some point we can break into the sort of monopoly of application layer. And we can say you don't really need to care about applications. All you need to care about is essentially foundation models and the Composio. And both of this combined sort of allows you to do whatever you were doing. Without even knowing at the back end of there is a Salesforce or there is some other CRM or anything like that and it's kind of already happening. Six months back this was not something people believed in. But now you probably would have seen a lot of your application interaction happens through ChatGPT or happens through other sort of chat interface.

Pablo Srugo (00:54:51):
You're positioning yourself as a key layer to this new language based UI. Where you know apps and icons, and stuff either go away or become no longer primary.

Soham Ganatra (00:55:01):
Exactly, and that's a massive position. That's a massive market, right? And so yeah, a lot of people were very bullish on that.

Pablo Srugo (00:55:12):
Perfect. Well, listen man. Let's stop it there. Let me ask the three questions that we always end on. The first one is, when was the moment when you felt like Composio had true product market fit?

Soham Ganatra (00:55:23):
Never, I still don't know if we do have a true product market fit. I think with AI, it's very difficult. The market is changing at such a rapid rate. It's very hard to say I actually have a PMF for a very long time. I would say there are moments in pockets. In the long duration when you feel like you have PMF. It's when a user comes and says, holy shit. I love you guys. Okay, maybe I do have PMF, right? Those are the moments, but in general. I would say on average, probably never.

Pablo Srugo (00:55:55):
And was there a time when you felt like Composio might just not work and things just kind of blow up, and fail?

Soham Ganatra (00:56:02):
Almost always, still like every day. It's like, yeah, because again with AI it's like, okay, you have a bunch of different questions. So you have questions like, do I really matter? Which is the existential question of, oh, foundation model will take over everything. AGI takes over everything, stuff like that. You have the other side of questions, which is like, hey, we are too small. What if we don't accelerate quickly and other companies take a bias, and sleep. Which is something also in AI. Which is happening a lot, right? So that would be another. A bunch of similar questions, I would say, but it's like, yeah, you're always afraid about these things.

Pablo Srugo (00:56:45):
And then last question, what will be your number one piece of advice to an early stage founder that's building today?

Soham Ganatra (00:56:50):
I don't know. Okay, so I can give you maybe one advice that is not obvious. But specifically for AI founders. Which is, don't believe too much in what customers are talking about. Basically, if you're building a new product and if you have very strong intuition on why it will work. I would say this is one of those times, where multiple products like iPhone could come out. Which is, if you go ask people. They won't really tell you like your product is what they need. But if you just go build it and show it, they would really get it. So for all the founders in early stages who are building in AI. Literally if you have very strong intuition, just go and do it. I think it's very easy to build. So it's definitely worth the whole set of effort and also there's a very good chance you are right.

Pablo Srugo (00:57:41):
Awesome, Soham thanks so much for jumping on the show, dude. It's been great.

Soham Ganatra (00:57:44):
I appreciate it. Awesome, thanks for having me.