Adit Abraham: The Unsexy Work Behind Reducto's $108M Raise

At a moment when artificial intelligence is increasingly defined by bigger models and louder promises, Adit Abraham and his co-founder Raunak are doing the work nobody talks about.

Before raising $108 million from investors like Andreessen Horowitz, Benchmark, and First Round, Reducto spent its earliest days doing things that didn’t look impressive from the outside: manually labeling documents, fixing edge cases no benchmark captured, and staying close to customers long after the demo ended. It’s precisely that unsexy work that allowed the company to build the infrastructure AI teams now rely on in production.

In this interview, we speak with Adit Abraham, co-founder and CEO of Reducto, about what it really takes to build an AI company that lasts, by focusing on what customers need now, proving value by showing rather than telling, and doing the work that actually moves the company forward.

We did a ton of manual, unsexy work in the early days. We tried hiring an initial data labeling team, but they weren’t accurate enough, so I spent a lot of my time just labeling boxes on documents. If you stacked all the pages that Reduct processed back when we were fully unautomated for Stripe billing and setup, it would actually be something like ten times the height of Mount Everest. Back when we were still fully unautomated for Stripe billing and setup, I would manually set up every single subscription.

Even though those things were repetitive and maybe boring, that was okay. Because what I cared about wasn’t whether I was doing the most glamorous work, but whether the company was moving forward. You’re lucky to be able to do that kind of work, because it means you’re signing up a new customer. It’s a privilege that you get to do that.

Ever since I was young, I always had side hobbies. In high school, I saw an article saying the creator of Flappy Bird was making $50,000 a day from ad revenue. So, my best friend and I in high school just immediately had this kind of reaction, forget school, forget all of that, we’ll just go make apps, and that's going to be our future. At some point, we even discussed not going to college.

Things didn't work out that way. We tried a few things, but did end up going to college and pursuing a longer career from there. But I think it was a really nice inspiration that kind of showed how going off the beaten path can lead to outlier outcomes for folks.

Even though we don’t work on game development today, I do think there’s something very valuable about seeing individual effort that sometimes just starts as a side project spiral into something much bigger. That eventually led me to go to MIT, where I did my undergraduate work in computer science.

I remember I was taking my first graduate-level ML course. It was a course on metalearning, like teaching models to learn. On the first day of the course, the professor introduced Raunak, who was a freshman in his first week on campus, and framed it as, "Hey, everyone, meet Raunak, he's going to walk you through how to do your first PSATs."

So Raunak was a learning assistant for a course that was primarily for PhDs. And that was crazy to me. Even though he had just come onto campus with really smart and exceptional people, he was already at the top. We became really close there, and the first time Raunak suggested that we could work on something together, I said yes immediately. I didn’t think twice about leaving my job or anything like that. He was just somebody that I admired enough for it to just be a no-brainer.

Before YC batch, we actually gave up on revenue multiple times. We tested different ideas, got to a point where people were willing to pay for it, but decided that the urgency with which they were willing to pay for it, or the need for which they wanted the product, wasn't high enough for us to want it.

And so, just to give you a sense of what this looked like tangibly, when we were selling Rememberall, we would constantly find that at best, people were willing to pay $50 or $100 a month.

So, Rememberall, as a product, was at that time the first long-term memory API for language models to remember things that you'd mentioned in the past. We would store context that was important and retrieve it when it was relevant. 

When we had demo Rememberall, we would find that people would get really excited about the fact that we were managing the files that they uploaded. We had put so much time into making that file management better. We started training our own models.

There are quite a few different ways that somebody can demonstrate how much your product means to them. It's not just the money, it's the time that they're willing to put into making the product great together.

We've put a ton of time into aggregating data. It's a big part of what we do, and it's a big part of why we've been able to train state-of-the-art models, but production data is different. We work with really intensive financial, healthcare, and insurance use cases that you're never going to find on the internet.

And so, really quickly, we started having customers who had tried public documents and saw exceptional performance, but they would come to us with the most esoteric examples imaginable. Like we've seen really hard cases where a doctor annotated things, and they just put things at the bottom of the page, and you were supposed to understand that it related to the thing at the top. We see really intensive financial tables with thousands of rows of data, everything along those lines.

The nice thing is our customers want us to solve those, and so we've always had this almost design partner-like relationship where they will come to us with that sort of feedback, and we will iterate day after day after day to make the models better.

And when you fix that feedback, they end up telling you whether or not that worked or it didn't. And you iterate by the end of that first week, you've already made a ton of progress with them. And that is really meaningful in that they care to make sure that your product is great. Like you're on the same team, you want to make the product better together because the work that we do directly helps them, too.

There's a world where we just rely on marketing. "Hey, it's the best product. Hey, it's state-of-the-art." All those things. But there are many companies that can say that. On the flip side, the other thing that we could do is actually put the product in front of people, even if it wasn't a perfect platform, to let them see on their hardest documents that it works to prove what you're saying is true.

And that translated to the company growing really quickly. At least in our case, being public in that way just meant that companies that otherwise probably would have ignored Reducto became really interested.

When we were a two-person company, a trillion-dollar enterprise decided to book a demo, and the reason they booked a demo was that we had public playgrounds where they uploaded hard documents that they'd seen fail on every other vendor.

And once they saw that work, that justified reaching out. If we hadn't done that, if we were this two-person company of 20-something year olds, I find it hard to imagine that they would even be interested in engaging with us. If we'd been shy about what we were building, we probably would have never gotten on the phone with them. 

I had known quite a few investors from just the course of building the company. And I think a lot of early-stage founders think in terms of firm brand. They only think of tier one VCs as the actual firm, which you know many of these firms have been around for decades, and have their own reputation from them.

But at the end of the day, the thing that matters most is that whoever you're raising money from, like that individual partner, is somebody that you're going to be partnering with for the next 10 years.

They're going to be there in all of your great successes, like your future fundraising rounds when you close the great contracts, but they'll also be there for the bad moments of the company. They'll be there when you have to let an employee go. They'll be there when you lose a contract. They'll be there when you have a big media incident, whatever could happen in the lifetime of the company.

But it's really important to see how their interactions changed when things weren't going well. I remember there was a moment where Liz, our seed investor, actually basically never takes time off. If I text her at 10:00 p.m., she's replying at 10 or 5. She's getting on the phone like doing whatever.

I think with every moment that is really exciting in a company, the thing that isn't discussed in interviews is what it took to get to that moment. When we were landing our first really big enterprise contract, it was an on-prem deployment, and we had never done an on-prem deployment before.

We didn't have infrastructure engineers on the team; we weren't this large org that could divvy up responsibilities. We would wake up, we would immediately go to the office, and we would be in the office until we were too exhausted to continue working. We would sleep for at most a few hours, and then we would go back, and we would try again and again and again.

People are diving in and doing anything at the company. There's no sort of notion of "Hey, if you're an engineer, you don't need to do customer support." There's no notion of "Hey, if you're an ops person, you don't need to label data for the ML team."

Because everybody just wants to see the company succeed. And the company succeeds when all of these things work. And when the product works, customers are happy.  And people don't think of their job in terms of whatever their role title is. They think of it in the capacity that they can help the company move forward.

So what I see reducto as it, it's not really just parsing. It's what does it mean to have this layer that connects human data to this new level of intelligence that applies across all of that data. We're seeing products built with what Reducto offers today that don't just read the documents; they actually create net new documents for their end customers. They do end-to-end work with agentic workflows.

In the future, most AI products will be some component of intelligence. That is what the foundation model companies provide, but there will be some components of context as well. And we want reducto to be the best way that you interact with that context, like a building block that you aggregate together and apply to a specific use case.