He fired all his customers. Then built a $1B startup in 2 years. | Jay Madheswaran, Co-Founder of Eve

Pablo Srugo (00:00:00) :
When do you decide to go all in on this product and shut down all your other business lines?

Jay Madheswaran (00:00:05) :
Pretty quickly. So we still had customers coming in for the old product, and we were getting a new pipeline going for the new product. And keep in mind, the product is barely there. And as a result, you have to be very careful qualifying the early customers to make sure they're aligned with the larger market. But what we noticed during that process was forty percent conversion rates from cold outreach into demo requests.

Pablo Srugo (00:00:28) :
Wow.

Jay Madheswaran (00:00:29) :
When we sent the, we're shutting down the service product, is when we got probably really strong signs of product market fit for that product. Because we saw other examples of cases where they were like, no, don't take it away from me. What do you need? That's always an option for people in the future. If they want to find product market fit, put it in the hands of people and then take it away, and see who complains the most.

Pablo Srugo (00:00:50) :
How fast do things go? How fast do you hit a million? How fast do you hit $10 million ARR?

Jay Madheswaran (00:00:54) :
The first quarter I think we hit a million in ARR, already for that product and the two months after that we had another million, and a month after that we had another million.

Pablo Srugo (00:01:01) :
You probably 10x'd in a year-ish from one to ten?

Jay Madheswaran (00:01:05) :
We've been growing a lot. I think the official number is eight hundred percent but I think it's ever-increasing.

Previous Guests (00:01:13) :
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 (00:01:26) :
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. Jay, welcome to the show, man.

Jay Madheswaran (00:01:42) :
Thank you, Pablo. Excited to be on, I loved watching your episodes.

Pablo Srugo (00:01:46) :
Thanks, man. So you started this business, I guess, five years ago. There was a bit of a pivot going on, but the last few years seem to have been completely explosive. I mean, you raised a Series A not long ago. Earlier this year from a16z, you raised $100 million at a $1 billion valuation a couple of months ago. So things seem to really be going up and into the right. So we'll talk about all of that, including the pivot, because I think what you start with and how you get to where you get to is just as interesting as what it is that you're doing today. So maybe as a first question, tell me a little bit about, you know, yourself, your background, and especially what you were doing in, 2018, 2019, 2020, right before you started Eve.

Jay Madheswaran (00:02:27) :
Yeah, absolutely. Thank you for asking me. So I mean, I'm an engineer by trade, largely. So I joined Facebook before the IPO, worked on a lot of their consumer facing products, and then joined Rubrik. Which I saw you had as co-host, the Glean founder, AJ, was my manager over there.

Pablo Srugo (00:02:44) :
Arvind from Glean was also here.

Jay Madheswaran (00:02:47) :
That's right, yes.

Pablo Srugo (00:02:48) :
Yeah, the whole Rubrik mafia, man.

Jay Madheswaran (00:02:50) :
Rubrik mafia is growing. It's pretty exciting, actually and that was a wild ride, right? Fom four people in a tiny room to five thousand and I grew to, at one point, lead about one third of engineering. 2017, 2018 is when I joined Lightspeed as a venture capitalist. I was leading their early stage AI investments back then. 2018, it's crazy to think about, back then AI does not mean what it used to mean today.

Pablo Srugo (00:03:18) :
That's, right.

Jay Madheswaran (00:03:19) :
But weirdly, a lot of the same type of issues. Where every company was saying they were an AI company.

Pablo Srugo (00:03:23) :
It was almost like the false start, the first wave of AI, pre gen AI. It was like the false start of AI where you talked about the things you wanted to do and most people couldn't actually do those things. And then all of a sudden, now you can.

Jay Madheswaran (00:03:35) :
Actually, that's exactly what led me to start the company. So what I noticed was, in 2018, one of the nice things about being at a platform like Lightspeed is you get to meet hundreds and hundreds of really exceptional talent. All at the top of their field, thinking about next gen ideas and what you start developing is a trend line of what's actually going on, right? Or at least what potential future could be and I think what I realized in 2018 is, you know, AI is probably there but this wasn't it. You know, a lot of the use case you were seeing were, supply chain, some sort of if then else statement claiming to be AI and the actual models to be trained to do even classification tasks weren't really good enough, right? But what was happening was literally year over year, they were getting good rapidly, especially in NLP and I started seeing signs of real use cases in NLP actually pop up. Largely from, you know, there's a lot of data labeling companies that end up doing a lot of training in NLP, but also some products started actually relying on the fact that transformer architecture and transfer learning with BERT allow you to go from needing a billion records to train a classifier model to literally one hundred the next year, to then ten, right? I mean, obviously with ChatGPT these days, you don't even need a training data set in large amount of cases if you want to, you know, qualify sentiment or something.

Pablo Srugo (00:04:57) :
Do you remember any mini ChatGPT moments? Obviously ChatGPT is where everybody saw the possibilities. It sounds like you're describing, well before ChatGPT, you seeing that things were on the come. I'm wondering if there were demos or products where you're like, oh wow, I can't, interesting that this is possible. Do you know what I mean? That we're kind of showing you a certain trend.

Jay Madheswaran (00:05:17) :
To me, it was something very, very simple that blew my mind. Funnily enough, it wasn't even ChatGPT like. But early on, one of the issues with computer vision machine learning is you would have to deal with, let's say, a large volume of legal contracts or bank statements and the typical way of extracting information from that used to be computer vision, right? So you train models, you get them to try to recognize pixels and boxes, basically and try to extract that out. And with BERT, entity extraction is becoming a thing. So a model could know if that's a name or not a name, is it a number or not a number, is it what have you, like a dog or not a dog and that alone made a dramatic improvement to extraction rates. So, right? So you went for needing to train thousands upon thousands of documents to ten documents, maybe, to extract things like dollar amounts from invoice totals, simple things like that. But it's happening in really, really large volume and the amount of effort required to train a model was just so cost prohibitive before it got reduced to tens of samples rate. That it just never made sense, even for really high volume repetitive grunt work and that's actually what got me to take a big leap of faith, and kind of bet on that trend line. And basically make it possible to help anyone make machine learning models easier. That's kind of what the initial talk track was for NLP specifically.

Pablo Srugo (00:06:46) :
Did you want to be a founder? You wanted one of those VCs that knew you wanted to operate? Because that transition, as much as VC and startups are close by. Obviously working hand in hand, it's not a simple transition. The life and the workday is very different. The risk reward is very different. How did you think through actually making that leap?

Jay Madheswaran (00:07:08) :
I think I'm still a builder, and I've always known that about myself. Even as an investor, I was actually building, you know, tools to help me prospect and identify resources, and doing a bit more data science-y work to help me do my job better. And I just can't help it, right? Forget building, I go make my own coffee table. Even stupid things like that, I still build on but part of my appeal to even joining Rubrik back in the time, Pablo, was I tried starting a startup within Facebook. Not that I had nothing else going on, but I tried doing that and I felt there were real gaps that I had in my own capabilities. And that was starting to get fleshed out at Rubrik. I just love technology and building in general so much that being a VC and investor allowed you to take multiple bets, right? And you get to kind of watch it from afar and still participate in the upside. And that, to me, was really intellectually stimulating. But what I can't get behind is the level of conviction. I was really, really, really high conviction towards AI, probably more so than it was rationally smart to do and given that conviction, I think that's when I knew this is not something I could just do by investing and focusing on a vertical to invest in. I wanted to do it, especially while I still had the operational chops to go do it and that's what ultimately lead me to making that jump. Plus, the team was a really, really strong team of people. I used to work together at Rubrik, was coming together, and I was excited to work with them again.

Pablo Srugo (00:08:42) :
What was the first idea in, 2020?

Jay Madheswaran (00:08:45) :
First idea? It's a good question. So the whole idea was around task automation at the time. So the idea was, think like the early day RPA bots were very error-prone.

Jay Madheswaran (00:08:56) :
Okay, I was going to say, yeah. RPA, yeah.

Jay Madheswaran (00:08:59) :
And the next level that we were kind of innovating on was, how do you make it kind of do an RPA-like bot automation on NLP tasks, right? That's how we got pulled into the legal industry, finance industry. We got pulled into multiple different industries. We were actually in multiple Fortune 500 companies. Deployed for those use cases , during our time.

Pablo Srugo (00:09:20) :
How do you get things off the ground? I assume being a VC was relatively easy to. I mean, you had the team from Rubrik, but then raising, you probably raised a pre-seed round. How did you kind of set things off?

Jay Madheswaran (00:09:30) :
That's right. So I think one of the nice things you do learn about being a VC is, how to actually fundraise in general. Because VCs are fundraising all the time themselves with LPs, and. 

Pablo Srugo (00:09:42) :
Yes.

Jay Madheswaran (00:09:43) :
You're hearing a lot of pitches every day. And I think that really helped me put together a very clear perspective on the market, and where we were going to participate. Along with a really robust set of companies in that phase, right? And that is a tactical advantage, at least when VCs go into it. They want some level of greed to be clicked on. It's like, oh, if they do this, I can see why they would make a lot of money and make me a lot of money. And you have to then, once you do that, reduce risk, right? Basically show a through line between where you are now to how you're going to get there and the more you can articulate it through a combination of storytelling, team, past experience, diligence in the market. All of these things are kind of tools to help you make that case.

Pablo Srugo (00:10:30) :
How much did you raise in that pre-seed round?

Jay Madheswaran (00:10:32) :
We raised $4 million initially.

Pablo Srugo (00:10:34) :
And this was when exactly in 2020, before or after COVID?

Jay Madheswaran (00:10:38) :
This was, funnily enough, the week of COVID. So the economy was shutting down.

Pablo Srugo (00:10:44) :
And you're raising four million bucks, yeah, nice.

Jay Madheswaran (00:10:46) :
And I mean, talk about founding stuff, right? So the first few years was pretty interesting and I remember our law firm didn't have a process. They relied on some documentation to be literally faxed at the time, and they couldn't go to their office. So some of the closing paperwork and filing paperwork for the state, so that we could get paid and get the EIN numbers, they all were delayed by a month or two. So we were having trouble paying ourselves and working for free the first few months.

Pablo Srugo (00:11:16) :
We're going to walk through that whole storyline, but tell me a little bit about, you know, you mentioned a few different industries. Maybe tell me a little bit about what you did in those early days. I mean, COVID happens. Who do you start working with? What are some of the specific use cases that you decide to tackle first?

Jay Madheswaran (00:11:31) :
We decided to go down high volume document processing. So that was kind of the initial starting point, which is CVE was pretty bad at doing these highly variant dog dives and what I mean by that are documents that have the same information, but look very different from one another. You know, think contracts, right? So contracts all have probably a name, a signing authority, and address, and email, whatnot. It has a few important pieces of information.

Pablo Srugo (00:12:01) :
It's crazy how much things have changed, by the way. This stuff is out of the box, post gen AI. Just put anything in, what's the name, it just knows, right? But literally in 2020, we're probably five years ago, it was a big task to use computer vision to say, here are two contracts, the names and dates and dollars are in different places, kind of like figure it out. I mean, that is not a simple problem to solve back then.

Jay Madheswaran (00:12:23) :
Exactly, and funnily enough, this is still not a solved problem. Which is the crazy part, right? So what if, maybe going to Eve, but one of the things we relied on is actually this architecture we built to continue to keep improving the models. That's where it gets harder, when you're chasing from ninety-nine percent accuracy to ninety-nine point nine. Gen AI still doesn't fully help with that. So you still have to kind of do more typical machine learning. But that groundwork helped us really set up well for making our Eve. Which is kind of this legal AI for plaintiffs, much better because getting the documents right is really, really important in legal. Trust is everything and OCR, silly things like OCR, information extraction, is still a really crucial problem. Especially when dealing with medical records, right? So if you get in an accident and go to a lawyer, they're going to pull medical records of where you got hurt, how much you owe in bills. So it's like all the old problems are back again. But it so happens that we spent many years building that tech, which kind of helped us out.

Pablo Srugo (00:13:28) :
So you're working with, which law firm you're working with and how do you structure it? Is it a design partnership thing, where they're sending you some contracts and you're pulling some information out? What exactly is that kind of first setup like?

Jay Madheswaran (00:13:40) :
The very first customer legal side was dealing with very particular workflows. We had a few in parallel, but one of them was doing case analysis. So maybe some precursor for people to understand plaintiffs work very differently than normal lawyers you're used to. These plaintiff attorneys are things like people that represent labor and employment, personal injury, and they don't really charge by the hour. They charge when they help you win a case, right? So if you.

Pablo Srugo (00:14:08) :
Contingency.

Jay Madheswaran (00:14:09) :
Contingency fees, exactly. That's what it's called and as a result, their business model is very different. So instead of retaining one client for longer, they're helping people that are in trouble, and they have to source them, qualify them, see if it's a problem or not, and then decide to take them on as customers. And then pour in the hours of legal work in hopes that they get paid. That's kind of the work. So if you think about it, what happens is, if the audience is familiar with a typical sales process. They have their own lead gen funnel to closing a customer type of playbook and at the top of the funnel, there's significant volume, right? So when you're talking about leads, there's sadly a lot of people that are looking for legal help that today don't get it. So there's a surplus of leads and as a lawyer, you just can't get on every single call, right? But with certain practices like labor and employment, you're technically at full employment. So if there's anything litigatable or you're owed in damages, things tend to happen at the edges, right? And this was established by the civil rights movement, and things from a long time ago are what we still rely on today. And you're looking for exceptions to the rule that are still protected by the law, and that takes legal expertise to do. And the challenge they have is they can't apply this legal expertise at the call center, right? So when you're initially calling them, it's just too expensive to put a lawyer on the phone to field those calls. That was the very first place where Eve was used. So how do you help them field this call? And now Eve actually literally talks to the client, which is kind of cool. It helps get them signed and qualifies them without violating any laws. You can't accidentally do unauthorized practice of law.

Pablo Srugo (00:15:56) :
Just to be clear, this is Eve, the post-gen AI product? Or even what you were talking about before?

Jay Madheswaran (00:16:00) :
No, this is post-gen AI.

Pablo Srugo (00:16:02) :
Okay.

Jay Madheswaran (00:16:02) :
The first product was literally just transcribed a phone call into a transcript, which itself was considered mind blowing tech by most.

Pablo Srugo (00:16:10) :
Yeah, when were you doing this?

Jay Madheswaran (00:16:12) :
This was 2023, like mid to late 2023.

Pablo Srugo (00:16:16) :
Okay, so at least post ChatGPT and even the transcription.

Pablo Srugo (00:16:19) :
Yeah.

Pablo Srugo (00:16:20) :
How do you walk into this? Tthis is the interesting part I'm trying to understand is there's a lot of B2B SaaS companies stuck in B2B SaaS that are trying to, everybody knows they have to add AI. So everybody's doing something, but very few are able to make that full transition where it's like, oh, you're truly an AI first company now. It seems like you guys have. So I'm trying to understand, before that October, November 2022, ChatGPT moment. Where are things at? What are you guys doing that allows you to then truly be AI first even in the second half of kind of your journey?

Jay Madheswaran (00:16:52) :
At 2022, and we were still a horizontal machine learning for NLP style startup with customers in legal, finance. Actually everywhere, supply chain and, you know, those were kind of the heavy verticals.

Pablo Srugo (00:17:07) :
And where were you at like, employee wise, revenue wise-ish. Just to get a sense of scale?

Jay Madheswaran (00:17:12) :
We were small. So we were about nine employees. I think a few million revenue, two or three.

Pablo Srugo (00:17:18) :
But you had a business. I mean, even that $2 or $3 million, it's a business. It's not like nothing's working. Might as well throw it all out, you've got a real sunk cost to worry about.

Jay Madheswaran (00:17:26) :
Oh, yes. It was possible to keep going. But what was becoming clear was you had to double down into a use case to derive more value, right? So you had to get the ACVs up higher, and that was a business led call. But to, you know, we're technologists, we also see where AI is going. We went from needing to train very complex NLP models to now you can just one shot things and what that made it possible for us was actually to go deeper into workflows. If you think about where pre-Eve we were used before, it was you would identify one very particular workflow that could be augmented by AI that's happening in very high volume, spend a lot of labor on the data science side, trying to get it to accurate enough that it makes sense to replace.

Pablo Srugo (00:18:13) :
What might be one, can you go through one workflow for me just so I can make it tangible?

Jay Madheswaran (00:18:17) :
So if you're doing large volume bankruptcy cases or credit accountants, right? What you get is, you get a document from the court saying so and so owes X money. And every single court has their own format for this. And then you have to go extract how much money, and to whom the money is owed. And then the law firm would then take that dollars and kind of go collect the money somehow, right? So they'll message people and collect it. That's a very concrete workflow and here there's a number of problems, where you have OCR issues, the actual amount placed might be in different places. You have exceptions to the rules always. Some judges will be like, this much so and so for attorney fees, so and so for whatever else fees, and then you have to add up the total yourself. So you have all of these different variations, and that was something we used to train models for. And to go deeper in the workflow just means, what is the actual thing they're trying to accomplish, right? And in the case of plaintiff warrants, it's to make a client happy, like they're wronged in some way. It's represent the clients, provide legal help to get them some sort of settlement. In civil litigation, it's almost always money as an outcome, right? You're getting some sort of payment for damages and that's, a workflow and process. And that's a workflow and process which is composed of about one hundred to two hundred different mini workflows that involve people, right? So what people are doing was the challenge is every single person is different, right? Every new person coming in, they have their own facts, they have their own problems, if you will and you have to take that, and convert that into those one hundred bespoke workflows. Which was economically unfeasible before LLMs. But with LLMs, we saw that we were actually able to go deep into the workflows and as a result, provide more value for the firms. And, you know, as a side effect, also get more value accrued for ourselves.

Pablo Srugo (00:20:13) :
So this is a junction in the story, basically an inflection point I want to go as deep as possible on. I think it's critical. You're at this point, you got $2 or $3 million ARR, let's say you've got nine employees, and you're horizontal, right? You're not just in legal, you're across finance, a few different verticals.

Jay Madheswaran (00:20:27) :
Yeah, we had to decide to throw it all away.

Pablo Srugo (00:20:29) :
This is exactly it, right? I find that, for what it's worth, this is the key. When you look at it from the outside in, everybody knows maybe the story of, say, Netflix. They used to do DVD rentals and they went to streaming, and you look at it ten years later, and you read about it. And you're like, wow, but also so obvious, of course they would do that. That's the thing to do and then when I talk to founders who have an existing business line, but they're not at scale. They're not $50 million, or rather $5 million in revenue, there is a genuine fear of letting all that go. Because you do worry about this shiny object type syndrome. You did that, right? So I'm just curious, obviously the tech piece, that I get. I mean, everybody would see, you know, what's possible with Gen AI and be like, but a lot of people might say, okay, let's stay horizontal, let's not burn our revenue, let's just find a way to augment with AI versus throw it all away. Walk me through, as much as you can, how you thought about that, what some of the internal discussions were, and how you ultimately made that pretty massive move. Which I think is a big reason for you being where you are today.

Jay Madheswaran (00:21:31) :
This is a tough call, right? I mean, I was up many nights, for months probably, before making this call. But what we did was kind of break down the decision into multiple steps. The only economically feasible way for us to continue growing to hundreds of millions of dollars in revenue was to go increase our ACV. So our sales productivity goes up, and be able to continue to hire.

Pablo Srugo (00:21:53) :
What was ACV before? What were you trying to get it to, do you have a sense?

Jay Madheswaran (00:21:56) :
ACV was tiny, it was like $6K or something, and we were trying to get it to $60K, right?

Pablo Srugo (00:22:02) :
Right.

Jay Madheswaran (00:22:02) :
So that requires doing something different, at least on the qualification side, product side, somewhere and that was kind of the journey we went on. And as a result, we just started working more deeply with customers to validate problems that they have. And we picked, you know, funnily enough, I think most people don't know this. But we actually launched a mini app for finance that went viral, very minorly, on Twitter.

Pablo Srugo (00:22:27) :
What did it do?

Jay Madheswaran (00:22:28) :
It was called Fin and what it did was, it was the early days of, you know, call it, you dump in an earnings statement of some kind. And you can then just do live analysis on it. And it was useful, for a short amount of time, people were actually using it a lot.

Pablo Srugo (00:22:43) :
This is one of the outputs of you talking to, in this case, finance customers, and you seeing that this could be a potential product?

Jay Madheswaran (00:22:50) :
The challenge you have is we have a timeline, right? So you don't have all the time in the world to make this call. We're trying to make a decision on this a month from when we decide that, oh, we got to go double down somewhere. We don't have that much time left.

Pablo Srugo (00:23:03) :
What was this time, by the way? Is this late 2022? What's the timeline at this point?

Jay Madheswaran (00:23:07) :
This is late 2022. This is probably December.

Pablo Srugo (00:23:09) :
So right, ChatGPT happened. You guys are like, oh my God, we got to do something. Let's explore.

Jay Madheswaran (00:23:13) :
It was independent of ChatGPT, right? It was more our own business was, we had to figure out a way to double down and then ChatGPT, of course, gave us a way to add more value. The technology made it possible.

Pablo Srugo (00:23:24) :
It's maybe an intersection of two things, but either way, you have to make a choice.

Jay Madheswaran (00:23:28) :
That's right. That made things, I guess, somewhat easier. It's like, you know, you have to focus and do something. But within that, because we were still spread broad and our revenue was kind of split in a few different verticals, we still had to do our work and do diligence on it. So what we basically did was try to go have conversations with our top clients in each one of those verticals and just spend time doing product discovery. On what does going deeper look like and basically working backwards without telling it to, like, how do we get to sixty thousand, one hundred thousand ACVs?

Pablo Srugo (00:24:00) :
How much do you pause everything to do product discovery when you don't have a product? Everybody does it. Everybody goes through this MVP motion. You don't have a product. You got to, you kind of got to do it, right? Once you have a business, you have, I mean, people are asking for features. You've got sales demos coming in. There's a kind of, nature of things are moving, it's hard to stop them. How hard do you go on just, you know, guys, we're not fixing bugs, we're not doing new features, we're just gonna talk to customers for a month? You know what I mean? How black and white do you do that?

Jay Madheswaran (00:24:26) :
It was not that black and white because, as you said, we still had customers, we were still getting new customers, we still needed help. Bug fixes were still important, we still were supporting customers that were paying us and recurring. As a result, honestly, it was myself and my co-founders just took care of ourselves to do extra credit work, right? So, instead of thinking about the future, about where this particular product goes, we instead went into what is the product direction, PMF area we want to really focus on, and that's where we spend all of our time.

Pablo Srugo (00:25:00) :
I'm going to ask you for a small favor, a tiny little favor. In fact, it's not even, now that I think about it, it's not even really a favor for me. I'm actually trying to help you do a favor for you. Just hit the follow button. You won't miss out on the next episode. You'll see everything that we release. If you don't want to listen to an episode, you just skip it. But at least you don't miss out. How did you set that up? What kind of questions were you asking? What were you trying to get out of your customers through these conversations?

Jay Madheswaran (00:25:26) :
Ultimately, we didn't hit the ground running with Ground Zero, right? So we already had relationships with a number of people in these firms and what we used to do was, they're also not silly. You know, oftentimes customers that adopt a startup are visionaries themselves. They're seeing around the corner. They understand something has happened and they also saw a ChatGPT independently of us. And as a result, they saw our technology, which was kind of doing what ChatGPT was doing in a way for them, right? But they saw that, oh, now you guys could probably also do this and this. So we had these, we called it product roadmap brainstorming questions. But, you know, we basically said, if we were to do more for you, that is business critical coming up this quarter, that you would drop everything and kind of get on and adopt it right now, what would it be? And that type of question, you want to really get urgency and get an understanding of what is the budget you're actually going after. Figure it out as early as possible and after that, you're honestly iterating on a week by week basis. There's no other way. You're looking at what's happening with each of those conversations, where is it moving, and applying your own sense of what the technology can do, and what the market can do into that. And when it became more clear, so when one of these workflows was figured out, so for example, let's take legal, right? So, in this case, the evaluation workflow was becoming clear to us that, you know, that has a lot of value.

Pablo Srugo (00:26:51) :
What's the evaluation workload exactly?

Jay Madheswaran (00:26:53) :
Case evaluation is when a new person calls a law firm and you have to figure out whether to take the case or not.

Pablo Srugo (00:26:58) :
This was one of the things that you built in response to the roadmap?

Jay Madheswaran (00:27:02) :
We haven't even built it at this point. So we had figured out through product discovery that that was a valuable pain point that people would be paying a lot more for. Now, it's a burning pain point and what we then did was basically, you have to validate it with the funnel. Because it's not only important for us that one customer does it, but we're able to set up a repeatable sales engine that is capable of continuing that sale. Like, is it actually a market, right? Not just a Hopia market and for that, honestly, it's back to hustle, right? I think I messaged all Amla one hundred and funnily enough, back then, we had eighty percent response rate to cold outreach. This is because they also saw ChatGPT coming.

Pablo Srugo (00:27:43) :
Yes, of course.

Jay Madheswaran (00:27:45) :
I think there's very few players with good resumes, if you will. That were reaching out to them and they actually just aggressively signed up, and got on calls. And we're not made for big law, so we're for plaintiffs and then we also had plaintiff firms coming. The interesting difference was the plaintiff firms came from network. Where the early people we were working with actually referred other friends of theirs and the big law largely came from outbound. And some connections, right? Most of these VCs are connected to someone in Big Law.

Pablo Srugo (00:28:13) :
Was the offer different? Was the product offer different for Big Law versus plaintiffs?

Jay Madheswaran (00:28:18) :
It was, yeah.

Pablo Srugo (00:28:19) :
What was the difference?

Jay Madheswaran (00:28:20) :
When we started digging into it with Big Law, it was about accelerating non-billable hours, largely, at the time. So non-billable hours, think about things a paralegal might be doing and the other challenge is it's incredibly bespoke. Every single Big Law firm has their own thing going on. So one, the culture and workflows are different. The second problem is they represent companies most often, and the companies have different things going on, right? And on top of that, they now have to follow ten, twenty different workflows for each one of those companies, even as a paralegal. So when we dug into it, it actually seemed like the amount of value AI was going to have for us at the time wasn't going to be that large. You're talking maybe five to ten percent improvements overall.

Pablo Srugo (00:29:06) :
For Big Law?

Jay Madheswaran (00:29:07) :
For Big Law and there was kind of this tension between not replacing billable hours, right? Because if you go after, if you try to make it too good for the company, suddenly the law firm is making it.

Pablo Srugo (00:29:20) :
This is the crazy thing about legal tech and almost like, one of the amazing pieces that's, you know, there's so many legal tech companies doing so well today. In spite of the fact that this has always been the problem with legal tech, which is if you make it too efficient and you just bring down billable  hours, you're just cutting down effectively the revenue. I don't know if now, three years post-Gen AI, they're just feeling like if they don't do it, they'll just die off, like they don't really have an option. But I totally understand, at that time, it was still probably facing a lot of resistance towards anything that would just, you know, where the efficiency is in non-billable hours, that's one thing. But if the efficiency just means you charge fewer hours because you do it faster, that's not so good.

Jay Madheswaran (00:29:58) :
Early on, most CIOs were excited, is kind of the initial thing. But they had to think about the business, you know, I think they see the value of what it can do eventually, but they don't know how and when to roll it out safely. Such that they don't cause damage to themselves and the damage is also multifold for defense firms. Because unlike plaintiffs, where you never have recurring customers, really, you don't want one person just suing everyone. On the defense side, you actually do retain customers. So the cost of a mistake is actually much higher. People tend to be a lot more conservative as well. But this is one of the hard decisions we had to make. Even though all this is true, there were still Big Law firms willing to spend hundreds of thousands of dollars to a million dollars with us on bespoke, almost consulting like engagements, right? At the time it was, hey, come customize ChatGPT for me, you know, make prompts and fine tune it a little bit for me. That was kind of what they wanted at the time. Versus for plaintiffs, we saw this consistent pattern about everyone's doing casework the same. They can't even Google it today. They don't even have processes for how do I make a demand letter? How do I make a timeline that is high grade and quality? All of this, they kind of figure out again and again across all these law firms. But they're all the same. It's all the same process and same workflow. There was significantly more repeatable, and they were seeing value literally, you know, when we made the product one weekend. Which I don't think we would have seen on the Big Law side.

Pablo Srugo (00:31:24) :
And for them, they don't have that conflict because they charge contingency. The faster they get something done, that's just money in their pocket.

Jay Madheswaran (00:31:29) :
That's exactly right. So our pricing and our business model is really aligned with theirs. And their business model is really aligned with the clients, right? Which is another thing people don't think about. But ultimately, when you need legal help, you want to find the best person for your task and as a consumer, you don't want to be on the hook for paying an unknown amount of money with no cap. Which is a scary feeling for anyone, right?

Pablo Srugo (00:31:52) :
One company I know in this space is EvenUp, and I know they've been around even before, pre-Gen AI. Where were they at when you were thinking about going? Because they're in kind of this demand letter generation. At least that's kind of how I understand the business, you'll know it better. But where were they, and how did you think about them when you decided to go after the plaintiff space?

Jay Madheswaran (00:32:12) :
Yeah, I mean, we were aware of them, but we didn't really focus on them. Because when we started, we were in labor and employment. Which is a very different market than personal injury. But what we did that was different was we kind of went all in on AI from day one. So from day one, we had built up a pretty robust eval stack and we wanted to do the case workflows. Versus even at the time was kind of done by offshore human labor on two very specific tasks and they're now layering in AI. And that was a pretty important difference. Because what ended up happening was labor and employment tends to go a lot more to litigation, which means you don't settle before you file a case with the court system. You oftentimes do have to file and then go into discovery before you settle. And that allowed us to actually work with our early customers on nailing the product across the entire workflow, which took a while, right? I mean, as you imagine, early on, if you have a chat interface, they can really type in anything at all. Which is not great from the engineering perspective. So figuring out the right product experience that ties that together with showing value quickly is what helped us kind of explode.

Pablo Srugo (00:33:23) :
And why do you go to labor employment versus bodily injury?

Jay Madheswaran (00:33:25) :
We do both now, but labor and employment was largely network driven. We thought it was interesting to go through the entire case lifecycle, and we saw a lot more examples of going deeper into the case lifecycle. Also, our early customers for labor and employment, so we just started doubling down into that vertical. And then basically, labor and employment firms, some of them actually do personal injury as well. And there's an overlap, and that became kind of the natural way for us to expand into personal injury.

Pablo Srugo (00:33:51) :
When do you decide to go all in on this product and shut down all your other business lines?

Jay Madheswaran (00:33:56) :
Pretty quickly. I think the scariest time was probably the first three months of making the call.

Pablo Srugo (00:34:02) :
And this would be what, early 2023?

Jay Madheswaran (00:34:03) :
Early 2023, correct. So we still had customers coming in for the old product, and we were getting a new pipeline going for the new product, right? And keep in mind, the product is barely there, you know, it was probably a ChatGPT wrapper at the time. Actually, it wasn't ChatGPT, it was Anthropic first, interestingly enough. But still, of course, we were doing more than just a wrapper but that's how it initially started and as a result, you have to be very careful qualifying the early customers to make sure they're aligned with the larger market, and the problems are roughly the same. And on the AI side, building evals is tough. You have to be careful in the space about, you can't really use client data to build evals. So you have to get clever with how you build it and as a result, we were kind of nose deep in both the customer discovery, the tech, and feasibility. Because you had to make a rough call on, is this even something we should do or not pretty early on and that process took about three month I would say. But what we noticed during that process was our conversion rates were ridiculous to interest in pilot. So we did the stage one discovery call, just to understand the need, but even before that, just outreach. So just outreach to customers and how many people sign up for demos or interest was significantly higher.

Pablo Srugo (00:35:20) :
Can you give me some numbers on that, whether it's the conversion rates, demos per day, et cetera?

Jay Madheswaran (00:35:24) :
You're talking forty percent conversion rates from cold outreach into demo requests, from one percent for the other product.

Pablo Srugo (00:35:32) :
From demo to pilot or demo to close, what did that look like?

Jay Madheswaran (00:35:35) :
Demo to pilot was ninety percent of the time.

Pablo Srugo (00:35:37) :
Wow.

Jay Madheswaran (00:35:38) :
Because, I mean at the time. When we showed them examples, it was interesting and what is unique in this space. Which we didn't realize at the time, was law firms weren't actually used to seeing working demos. You know, it sounds silly, but they're used to buying services, right? So you oftentimes pay someone by the hour to go write a demand for you and you don't really know until you get the demand back if it's worth consuming or not. And this whole idea of a demo kind of blew their minds.

Pablo Srugo (00:36:06) :
What was your demo, actually? How did it work?

Jay Madheswaran (00:36:08) :
We basically fine tuned it for. Not fine tuned it, but made the product really, really good at one workflow we knew was good. In this case, case evaluation, and then soon demands, and then brainstorming use cases came right after that.

Pablo Srugo (00:36:23) :
That demo would be what? You'd get them on a Zoom or whatever, and you'd upload, or they would upload some new case they got, and you'd do the evaluation right in front of them, or how?

Jay Madheswaran (00:36:31) :
Of course, yeah. I mean, we went and downloaded some documents from Kspacer, which is a public data set of cases, and had some client transcripts and depositions in there. And showed them, what do you want to learn about this? Let's make an evaluation of where this case is at and you just type that in and their seeing it live, right? And that kind of blows their mind, and then you set expectations. So then after that, for pilot, at least early on when the product was still iterating, you set clear expectations. You don't trick them, you tell them, hey, listen, it's early. It's going to iterate, but you get this amazing team to help you through it.

Pablo Srugo (00:37:03) :
And if you think about it in terms of ROI, how much were you charging relative to what they were paying? Was the delta big or was the time big? What was the biggest, besides the fact that it was impressive technically. What was the underlying ROI for them?

Jay Madheswaran (00:37:21) :
Case capacity, is something they look at, which is how many cases can an attorney handle it once and that is indirectly tied to how many cases their marketing engine can feed them with. How many cases are they able to kind of move on to the next step so they're not actively working on it? And ultimately, all of that is labor constrained, right? So they're working on tasks. You know, one case might be stuck at go interview the witnesses, right? And one might be stuck at respond to the opposing counsel's request for discovery. And they're all very different tasks, but they all have legal work involved.

Pablo Srugo (00:37:53) :
But your first product, which part of that funnel did it deeply accelerate?

Jay Madheswaran (00:37:56) :
The top, the intake. So that was directly tied to now you don't need to be limited by case managers and call center people. You can accelerate the number of cases and also the cases that qualified because you could run through that at least first pass of lawyer like intelligence early on.

Pablo Srugo (00:38:13) :
And you would feed it, like somebody would call in, they would still presumably at that point talk to a human. But then that conversation would be fed to your AI and then it would tell them, yeah, this is a case we're taking, this is not a work case we're taking, kind of. 

Jay Madheswaran (00:38:24) :
Yeah.

Pablo Srugo (00:38:25) :
Is the output.

Jay Madheswaran (00:38:26) :
It would give them, I mean, obviously it's not going to say yes or no.

Pablo Srugo (00:38:28) :
But it's enough qualitative output so that they could easily make a decision on, and that's for them. I mean, that basically. Again, like you said, they can just review so many more cases and therefore take more cases on ultimately.

Jay Madheswaran (00:38:40) :
Correct, yep and I can give you an example of why stuff like this is annoying for lawyers. In one case, if you have missed wages as the main claim, meaning an employer, when they fired you, didn't actually pay you properly. For example, they didn't count lunch breaks right. They didn't count check in, check out times right. They kind of chipped you half an hour here and there. That you get back in a worksheet format or sometimes clients literally just output calendars in screenshots. That's the type of data law firms are dealing with, right? And to do case evaluation properly, ideally someone went through, tallied up how many hours you are missing potentially, and what's your hourly rate, and how much dollars are you missing? That's what ideally you did, right? But instead, because that's so time consuming, a lot of times the call center reps were trained to just say, is there a missed wage? Yes, then they would pass it on, right? But that could be a bad case and you end up wasting hundreds of hours representing them when it probably wasn't worth it for anyone involved.

Pablo Srugo (00:39:44) :
And once you go through these three months, you're seeing these insane numbers. Do you literally call your non plaintiff customers and just churn them? Or how do you actually effectuate that? How do you make it happen?

Jay Madheswaran (00:39:56) :
You know, the engineering team did an amazing job of managing that, actually. So the product was low touch enough that largely what we were on the hook for was bug reports. So we pretty much communicated, hey, we're going to stop development on new features. If you want to keep renewing, we'll allow it to happen for this amount of time and if not, no biggie, we'll kind of eat the costs. Let us know, in some cases where it was feasible to do. So we were kind of generous with how we let people go and interestingly, maybe a surprise learning during that. When we sent the, we're shutting down the service product, is when we got probably really strong signs of product market fit for that product. Because we saw other examples of cases where they were like, no, don't take it away from me, what do you need? So that's always an option for people in the future. Want to find product where it could fit, get it in the hands of people, and then take it away and see who complains the most. Not a bad way to go.

Pablo Srugo (00:40:57) :
How long do you kind of stay in this beta pilot type phase? Where do you get the product to until you decide to say, okay, this is working, let's make it happen?

Jay Madheswaran (00:41:06) :
A few ones, so we had a few kind of 100k plus deals come in pretty early on. As soon as we started the firm and something magical happened where once we had the initial workflows and the additional workflows they wanted working. Usage on the product went up by itself without us doing anything and what I mean by that is, we had kind of put restrictions to stop them from inviting other people. This for stupid reasons, we ran into QPS limits on Anthropic and OpenAi, and scaling issues at the time. They were also scaling and as a result, we put limitations on who could invite users into their account. But they found a workaround, which was kind of a bug on the product, and they started inviting their entire company on it when we had planned a three month rollout. And within a month, they had just kind of onboarded everyone else themselves. And on top of that, we started seeing usage every day. And then within a month, we started seeing signs of long sessions. So we track how long is the average session on Eve and we started seeing multiple four hour plus sessions of people. I mean, who knows what they were doing there, but they were doing some deep work continuously for four hours. Which was kind of scary at the time, right? Because AI back then wasn't good at that many turns. So we were like, oh, I hope they were doing this right. So all of this showed that there was strong interest in using the product and in a period of checkpoints we had with the managing partners of these firms, they were seeing value. You know, they were giving real anecdotal evidence around, Oh, I took this case on that I would have missed completely otherwise. Oh, this person sent a discovery request the same day, which went over taking a month to do so before and these were kind of the anecdotal pieces of information we were getting. And that enabled us to basically make a hard call. Okay, this is it. Let's just go for it.

Pablo Srugo (00:42:56) :
And when you went for it, tell me a little bit about go to market. I think is something that every founder wants to know more about. What were some of the go to market tactics you used that either really worked or didn't work at all?

Jay Madheswaran (00:43:06) :
You know, early on, like I mentioned, cold outreach worked surprisingly well, better than expected.

Pablo Srugo (00:43:11) :
It's cold email campaigns?

Jay Madheswaran (00:43:13) :
Yeah, cold email campaigns and actually calling them. Because every software firm has a phone number publicly available, but they gatekeep a lot because everyone calls them. But if you sound interesting enough, they will let you through. The other one is LinkedIn sometimes works for people that are managing partners and tech oriented. So that was another channel, but eventually, once we had the first few customers, we really leveraged our network, right? So really blowing it out of the water for them such that they'd want to talk about it naturally without us forcing them.

Pablo Srugo (00:43:43) :
Yeah, I was gonna say, did you run a referral program or was just organic word of mouth?

Jay Madheswaran (00:43:46) :
It was organic word of mouth but we asked them, where do you go, right? I mean, where do you find more of you?

Pablo Srugo (00:43:52) :
And are they not competitors? Maybe it's a dumb question, but why do they tell others?

Jay Madheswaran (00:43:56) :
So this is actually an interesting question. So they are competitors, but they're competitors locally, right? So if you look at it, you know, in, I don't know, Los Angeles, in one neighborhood, there might be two or three personal injury, labor and employment attorneys representing that area. They'd be competing strongly with each other. But outside the city, outside the state, they all have their own friends that are teaching them tips and tricks. Because similar to what I said, there's no book on this, right? If you go online and say how do I start a plaintiff firm. There aren't very many good materials, and a lot of things are on the edges. How do you know which cases to evaluate? Even that is acquired knowledge for these law firms, and they actually share that information. The second thing that happens in this space is there's incredible wealth disparity between the plaintiff side and defense side. Where defense, because they're paid by companies, they have massive paralegal teams. When they show up to court, they literally show up with ten people. On the plaintiff side, you have one or two people there, right? And they're used to being kind of smothered like that. Where defense, one tactic they have is actually threatening trial work and litigation work. Which takes up a lot more time for the plaintiffs. Which means the plaintiff has to decide between, do I do this trial work or settle these smaller cases just to pay my family, right? Get paid and that is a real pressure that happens. And with AI, that has completely been flipped on the head, right? With Eve, Eve now is incredible at helping individuals with these workflows and as a result, you're just putting pressure back onto the defense side, and they love this. And some of the ways they were doing this before technology was they go to conferences and share, tips and tricks. Like, hey, in a closed room, right? Here's how I defeated State Farm's defense team in this particular litigation, right? And here's who you should kind of settle with quickly, here's who you should kind of push to trial because they don't know what they're doing in trial. And all of these types of tips and tricks don't exist anywhere, but the only way to share it was by meeting people locally and getting to know them. And because the product was so good, they were actually bringing us naturally into these events. And then we kind of doubled down and made that into an actual motion.

Pablo Srugo (00:45:59) :
First full year 2024, how fast do things go? How fast do you hit a million? How fast do you hit $10 million ARR?

Jay Madheswaran (00:46:05) :
I think from when we GA'd the first quarter. We GA'd probably end of 2023, probably early January. The first quarter, I think we hit a million in ARR already for that product and I think the two months after that, we had another million. And a month after that, we had another million. So it's kind of been rapid growth since.

Pablo Srugo (00:46:25) :
You're probably 10x'd in a year-ish from one to ten.

Jay Madheswaran (00:46:28) :
We've been growing a lot. I think the official number is eight hundred percent but I think it's ever increasing.

Pablo Srugo (00:46:34) :
And tell me a little bit, maybe just high level. What are some of the things that Eve does today? At first it was just that first piece, right? Just the evaluation of kind of a qualitative piece of what this case is about or whether you should take it or not. How much have you penetrated these organizations today in terms of what you're doing for them?

Jay Madheswaran (00:46:48) :
So it's going deeper and deeper into the casework. But if you were to up level what Eve really helps with, it's ultimately AI as an efficiency thing, you know, pretty much everywhere. But how do law firms actually get value out of it, right? And this is a combination of a few different factors, like you need to get the product right, you need to get arranging the people to use the product right, and the technology has to be really good, right? To make all of this happen and we call this AI transformation. I think we were kind of in this whole trend of digital transformation for a long time. Now we're in this workflow where work that was done by people can be done by software. But all that's going to do is change what people have to do and in anecdotal evidence so far, every firm that has us has chosen to do more with their people than let them go.

Pablo Srugo (00:47:31) :
But are you like answering calls now? You're doing voice, like you're doing kind of full AI features?

Jay Madheswaran (00:47:36) :
Exactly, so part of the benefit is we've kind of taken the best of the cultures from my time at Facebook and my time at Rubrik. And built a culture of high engineering discipline, launching high quality products that work for the most secure of environments. And the Facebook speed, which we operate also, as you can imagine, in a fully AI driven way. You know, we do AI prototyping to do PRDs, and we iterate quickly. Of course, we use all the live coding tools and that's enabled us to launch incredible features in weeks of effort, right? And on top of that, we're also changing our tech stack, whatever makes sense. Before, if you imagine a year or two ago, you had to put a lot of effort into very bespoke rack stacks and routing, but now LLMs are getting more capable of doing more agentic behavior. As a result, you have to change your architecture. Otherwise, you're going to be left behind. But that inefficient motion has helped us go deeper into workflows that were otherwise building our industries by themselves, right? So, we talked about voice. So Eve now has a voice AI component where it's actually able to take the call and do this live filtering of is this a good case? It can even be empathetic, right? Which has been trained on personal injury cases and labor and employment cases and on top of that, deeper into the workflows, like getting this medical chronology demands right through pure AI is really hard still. Because there's very low tolerance for errors. You know, if you get a case wrong, you're just going to lose that settlement and the cost is too high. So, we have put a lot of work into making that accuracy as high as possible and that ends up being the source of truth data, which then helps all the other use cases.

Pablo Srugo (00:49:11) :
Perfect. Well, Jay, let me stop it there. I'll ask the last three quick questions that we always end on. The first one is, when did you feel like you'd found true product market fit?

Jay Madheswaran (00:49:18) :
I think that moment where I mentioned where usage, customers were working. Breaking our system to go invite people, to go from like one user to, I think, a hundred attorneys getting on board in four weeks. That was like a, oh my God moment for me, for sure.

Pablo Srugo (00:49:32) :
And was there a time where you thought things might just not work out?

Jay Madheswaran (00:49:36) :
I think one of the difficulties of being a founder is you need to have full conviction in whatever you're doing, and unbridled enthusiasm. I think I do have that, I'm hard enough to know that it's reality. You have to react to data that you actually see, but when you're chasing something that you have data points for, you have to see it through. You know, unlike a VC where you take multiple bets, you have a portfolio of one and you have to make it work. So were there doubts in individual execution? Yeah, absolutely. But were there doubts overall that we would succeed? I think no.

Pablo Srugo (00:50:08) :
And last question, what will be your number one piece of advice for an early stage founder that's looking for product market fit?

Jay Madheswaran (00:50:14) :
I think it's to set really high expectations of where product market fit is, right? I think early on, so before you make your first million, you have to really focus on the stage all the way from the top of the funnel. So meaning, how are you doing your outbound activities? How quickly are they converting? And even that is a pretty good signal early on, because that tells you how close to home you are in urgency of problems and how articulate you are with what you're actually solving for. And I would track those percentages, right? So when you're doing something, if you're really early on, you are iterating on multiple different ideas simultaneously. So at that point, you just literally track. I emailed a thousand people, two people got back to me on this campaign. I emailed a thousand people here and fifty got back to me. Directionally, that's better but you should aim for something even higher.

Pablo Srugo (00:51:01) :
Well, just to put it into numbers. I mean, forty percent demo means you email one thousand people, you get four hundred demos, which is insane. It's almost too much, what do you do with that, right? But that just shows you just how much people resonate.

Jay Madheswaran (00:51:14) :
That's right, that was the problem. We didn't have capacity to actually handle those calls and that was a good problem to have back then. And I think we still have it, but that's kind of the bar you should shoot for. I think the second thing is, when you do hit product market fit, this is like, call it a million plus. When you have real product market fit, it just feels different. You know, I hate to say this, but the Paul Graham essay, you know, the Peter Thiel essay, like everything that all these people have written is kind of true. Where every single metric you do track suddenly looks different and that is how I think you know for sure you hit product market fit.

Pablo Srugo (00:51:48) :
Awesome. Jay, thanks so much for sharing your story, man. This has been great.

Jay Madheswaran (00:51:50) :
Fantastic. Thank you, Pablo, for having me on.

Pablo Srugo (00:51:52) :
Wow, what an episode. You're probably in awe. You're in absolute shock. You're like, that helped me so much. So guess what? Now it's your turn to help someone else. Share the episode in the WhatsApp group you have with founders. Share it on that Slack channel. Send it to your founder friends and help them out. Trust me, they will love you for it.