If you’re a Reducto customer, you can text or Slack founders Adit Abraham and Raunak Chowdhuri at any time.
It’s a simple, but effective instrument for keeping a pulse on customers, as close to sidling up to their desk as an early-stage startup can get. Abraham credits it as one of the most powerful tools in the young startup’s toolbelt. But it’s, admittedly, not the sort of splashy advice you typically look for when trying to parse a breakthrough company’s path to product-market fit — and that’s precisely the point.
“There’s a disparity when founding a company. We hear inspirational stories of these grand visions and the big-picture ideas. So when you’re a new founder and you’re thinking about how to talk to customers, it doesn’t feel instinctively right to just reach out to people on Slack. It feels like you should be doing something flashier or cooler,” he says. “But in the beginning, it only matters whether or not people are using your products.”
He likens novice founders to newbie exercise enthusiasts — they both tend to overcomplicate things:
If you ask a bodybuilder for advice to get fit, they would tell you very simple things: Focus on your nutrition, go to the gym, prioritize sleep. But you see folks trying to get into fitness will instead agonize over figuring out the perfect training regimen and complex routines.
Most company profiles chronicle the rise to success long after it’s already been cemented. While we’ve shared a number of stories in our “Paths to PMF” series featuring companies that have long since found extreme product-market fit, we think there’s equal value in mining the stories of startups that are a few clicks earlier in their PMF journeys. As time removes you further away from the early days, recollections get hazy, and tactical steps slowly morph into broad platitudes like “stay close to your customers.”
But before we dive into the nitty-gritty of how they made it all happen, a brief explainer of the problem that Reducto is solving: 80% of the world’s business data is in unstructured file formats (like PDFs or Excel spreadsheets). If you’re trying to build a reliable AI application for handling, say, insurance claims, health records or financial statements, you’ll find that even the best LLMs start to hallucinate trying to read these unstructured documents. Multi-column layouts get jumbled together, key figures are ignored, and tables are a nightmare. Reducto reads these complex docs the way humans do and creates LLM-ready inputs with remarkable accuracy.
And while Reducto is still early, the trajectory has been remarkable, and as seed investors First Round has been ringside as Abraham and co-founder Chowdhuri hurtle past milestones. In just one year, they went from a pre-seed to a seed round to their just-announced $24.5M Series A and nabbed a Fortune 10 enterprise customer. If you printed out every page that Reducto has parsed so far, it would be something like 3.5x the size of Mount Everest.
In this exclusive interview, we pluck out some of Abraham’s biggest lessons on building Reducto so far, while they’re still incredibly fresh. And while there’s advice that follows specifically for founders building in AI, other early startup folks will have plenty to jot down in their notebooks — from Abraham’s one-man GTM show, to how they’re keeping the team quite lean. Let’s dive in.
Early brushes with entrepreneurship
Abraham’s entrepreneurial ambitions can be traced back to perhaps an unlikely place — the mobile game Flappy Bird. “Games were the thing that drew me to computers and Flappy Bird was what made me interested in programming,” he says. So as young, ambitious high schoolers, Abraham and his friend started making their own apps. The rumors that Flappy Bird’s creator was netting $50,000 a day off the app didn’t hurt, Abraham laughs. “We had these moments of like, ‘Forget school, we should just be making apps and we won’t even need to go to college.’”
So the duo spun up a goal-setting app, and while it didn’t make Abraham a newly minted app mogul, it did launch something of a side hustle. “I learned more about marketing than I did about building apps. We started making Instagram pages to promote the app, where we would post summaries of books like ‘How to Win Friends and Influence People’ and those pages started taking off. Even when the company shut down, I started running Instagram pages for other companies,” he says.
And while the goal-setting app might have fizzled out, Abraham’s entrepreneurial spark was just starting to ignite. He went on to M.I.T., where he eventually joined the Accelerator program and started building Sidewalk, a Shopify plugin that would recommend products from other stores during the checkout process.
But it was a chance classroom encounter that would truly turn the tides for Abraham’s founder path.
The team up
Abraham, then a junior at M.I.T., had walked into the first day of a graduate-level ML course with a heavy dose of undergrad imposter syndrome. So imagine his surprise when a freshman was standing at the front of the lecture hall. “The professor introduced Raunak to the class and told us he would be walking us through the first problem set. It was kind of obscene to be sitting there and see this person who had been doing ML research since he was 12,” Abraham recalls (Chowdhuri actually had 100 citations to his name before he even got to campus).
The two struck up a friendship, joined the same living group at M.I.T., and stayed in touch after Abraham graduated. After running Sidewalk for a year, he joined the ads team at YouTube. And frankly, he was bored. “We were both in a transition phase. Raunak had wrapped up his previous company and was going to be graduating from M.I.T. soon. I was vocal about being ready for something new,” Abraham says. So they did what plenty of other young engineers do — they started applying to hackathons together (they actually won Anthropic’s hackathon fresh off the release of Claude).
After a few of these hackathons (and a few trophies) under their belts, they finally broached the topic: Did they want to start a company together? “At least for me, it was an immediate yes. There was no question in my mind that Raunak would be an exceptional person to work with,” Abraham says. Their areas of interest also tipped the scales — Abraham was more drawn to the GTM and product side of the business, Chowdhuri was deeply technical.
At the time, I wasn’t focused on the idea that we would work on or whether we could get funding. I was just excited to build with this person.
So they plucked an idea from a past hackathon project Chowdhuri had worked on, and applied to YC, getting accepted into the Winter ‘24 batch.
The initial idea
Their initial YC application was far removed from PDF processing — it was a long-term memory solution for language models. And the co-founding duo gave themselves a simple rule: “We decided if we got into YC, we would at least stick with the idea for the duration of the batch. We were worried about just pivoting around aimlessly,” Abraham says.
But there were some early signals that, while attention-grabbing, their idea didn’t have legs. “It did go somewhat viral on X — people thought it was cool to see a language model bring up something you mentioned in the past,” he says.
Hundreds of folks asked to be onboarded to the product, then called Remembrall (“Harry Potter” fanatics will notice a pattern with these company names), but early onboarding calls hinted that they were a bit too early to the space.
They had conversations with enthusiasts who would maybe try adding it to their AI application. “It was never the case where people were saying, ‘I’ve noticed that customers get angry at my product because of X, Y, and Z and Remembrall can fix it.’ That was the first signal. The second signal was that people were maybe willing to pay $10-20 a month because they found the idea interesting. But no one’s product team needed it,” he says.
We set up demo calls with everyone who expressed interest and we asked what their use case was and what problems the product would solve for them. And we would very rarely get a real answer. It was more of a ‘This seems cool’ curiosity.
But buried within these somewhat directionless enthusiast chats was an ember of an idea that needed fanning.
The pivot
“One of the most common feature requests for Remembrall was, ‘Hey, you’re managing my user’s chat history, can you manage the files that they’re uploading as well?” Abraham says. “We went into our projects expecting to help teams with fun ML problems, but very quickly learned that one of the biggest bottlenecks across most pipelines was actually well before retrieval or generation.”
So the co-founders built a very simple solution — they would upload the docs, parse them using an external parsing solution, and then chunk the information for you. “It’s embarrassing to look back on it now, it was a very ugly Streamlit app with a super simple document segmentation tool. It would do nothing other than split your documents into little boxes of content. I can’t stress this enough — it was literally a weekend project that we threw together,” says Abraham.
But they posted it on YC’s forum anyway, and the response was immediate and overwhelming. “We started getting replies, ‘This is better than what I’m getting from Textract. Is this a hosted API? Where’s the Stripe link?’ The pull was much stronger than everything else we had worked on,” he says.
“We knew how annoying it was to build a PDF processing pipeline. We knew how much time we were putting into post-processing that content. So when other people mentioned the same painful experience, it was a clear aha moment,” he says.
Unlike the “seems cool!” enthusiasm for extended LLM memory, there was a distinct need and a massive pain point. “These companies are not trying to spend many hours of engineering time on PDF processing, but it’s the bottleneck that’s stopping them from building the things that actually matter. If we can be that ingestion team for our customers and take that problem off their plate, that’s clearly very valuable,” he says.
We had spent enough time exploring things that were not resonating. In comparison, it felt like we were getting punched in the face by this new idea — it was something people cared about an order of magnitude more.
The competition conundrum
But let’s be clear here, while the team was getting inundated with signals that they had sniffed out a winning idea, they were somewhat hesitant to wholly embrace this new direction. For one thing, neither co-founder had expected to build a PDF processing company. And there was an initial gut fear that they wouldn’t be able to build a moat quickly enough.
“PDF processing is not a new space,” says Abraham. (Brace yourselves here — PDFs have been around longer than both Reducto founders have been alive.) “Every day we were looking into this idea, we’d find a new, different solution that claims to solve the problem. But the way we thought about it was that everything on the market will get you part of the way there. Nothing we tried was at the accuracy level for what we wanted, and clearly not at the level our customers wanted.”
Here’s how he explains the difference in layman’s terms: “There was an old era of parsing PDFs using the metadata and just trying to read the file itself. Our insight was that these documents were made for humans, not machines, to read. Every little visual cue, like a gap between two paragraphs, is me telling you, ‘Hey, this is a new semantic piece of information.’ Or a tab structure in a list tells you, ‘This is a sub-idea of that parent idea,’” Abraham says. “We wanted to read these documents like a human would. And that is a very deep problem, but thinking about it from a first principles lens unlocks a whole range of different ways to improve parsing.”
I think younger founders like ourselves are more afraid of competition than they should be. Our question wasn’t, ‘Is the market big enough?’ The question was, ‘Can we be better in a meaningful way?’
Once the founders felt confident that the answer was yes, they put blinders on. “We didn’t want to spend a year or two of our lives parading around pretending to build a company. We just wanted to focus on this one thing,” Abraham says.
The MVP
Today, Reducto has six different vision models to ensure accurate outputs with any document, but they started with just one. “We trained the first model for layouts, breaking down a document into sub-regions. As a human, when you look at a multi-column layout, you know how to read it. But that’s not immediately obvious if you have text just sitting there. You’ll read the text from left to right, you might merge the columns together, etc. And so we tried to break it down into how we understand the paragraph structure and the position,” he says. “We actually see PDFs as just images, and so we convert them to images. It’s not about PDF as a standard, it’s about documents and human content overall.”
They cobbled together an initial pipeline and paid UI barely any thought at all. They managed to officially launch just two weeks after that initial post on YC’s internal forum — although they were hesitant to do so, despite prodding from folks to get the tool out there.
We got a lot of advice to launch early and not be afraid of failure. But we knew we were in a space where if we launched too early and the product wasn’t good, it would just be a waste of everyone’s time.
“People already had PDF processing solutions, Reducto needed to be decidedly better. Then once we cleared that bar, we would try to launch as soon as possible,” says Abraham.
They kept the launch simple. “We just put up a post on our social channels saying essentially, ‘Hey, we know PDFs suck. Reducto’s building vision models to address that. Try it out for yourself,” says Abraham.
But here’s the genius behind their launch buzz: “We had a playground with a pre-populated document, but you could also upload your own file. People usually had some sort of file that they’d seen fail with parsing all the time. Seeing it work with Reducto immediately sparked attention.”
Inbound interest poured in from startups, consultancies, even BigCos (by the time demo day rolled around in April, Reducto was getting 55K in monthly traffic). But Abraham and Chowdhuri were intentionally choosy about who to initially partner with.
The first customers
To sidestep the enthusiast problem that they had faced with Remembrall, the co-founders wanted to focus their attention on an ICP that felt the pain most acutely. That meant industries where accuracy was imperative — like finance, legal, or insurance. Making an error on someone’s insurance claim isn’t just a silly typo.
And while they started having some initial conversations with enterprises that had expressed interest, they mainly targeted startups. “They would be able to integrate us by the end of the week,” Abraham explains.
The onboarding experience was white-glove, and even today, Reducto does not have fully self-service onboarding. “We still manually onboard every customer because we learn a lot from the process,” he says.
It was also a quality control mechanism. “We didn’t want to be in a situation where someone integrates us into their production environment and all of a sudden our infra wasn’t ready to handle the next million pages we were hit with. We wanted a failsafe on the customer experience,” Abraham says.
Pricing for acute pain
To sketch out their pricing model, Abraham balked at the more expected usage-based approach and instead went with a tiered system that starts at $300/month. “Our pricing tiers assume that you’re processing at a certain volume. The startups that we work with tend to be processing above 15,000 pages a month, at minimum. In some cases, hundreds of thousands or even millions,” he says.
Part of this thinking stems from a bit of scar tissue from the initial enthusiast reaction to Remembrall — and the $10/month price tag that came with it. “We want this to be something that’s tangibly impactful and real infrastructure for companies to build AI products. In order to do that, we wanted use cases that matter to the end customer.”
Enterprise comes calling
While their preliminary focus was on getting quick-acting startups onboarded into the tool, it didn’t take long before enterprises took notice of Reducto — with one seemingly landing in their lap just a few days after publicly launching. “A Fortune 10 company actually tried out the playground and signed up for a demo call,” Abraham says.
After some exploratory calls, at one point during the sales process they brought 14 of their engineers to meet with the Reducto co-founders for an entire day — the signals of developing PMF can’t get much clearer than that.
“Their entire team just sat with us to learn more. The depth at which they cared and had thought about the details showed that it mattered to them. I didn’t feel that we had product-market fit until our enterprise conversations started getting escalated in a way we didn’t expect,” says Abraham. Eventually (after 154 days total, over 20 hours worth of meetings, an acquisition attempt and a few hundred emails), the Fortune 10 signed on.
But other enterprise deals were an even slower burn. Abraham’s former college roommate was on the team at Scale AI and connected them to Reducto, but it would take another year before they eventually signed the deal. Here’s Abraham’s advice for maintaining focus — and patience — during these prolonged sales cycles.
“The clearest signal is how much your champion seems to care. There are enterprise deals where I’m reaching out to the champion almost begging them to get on a call and they never respond to me. Other times, there are just hoops you have to jump through, maybe their legal team is taking forever, but your champion still seems very engaged. You can clearly tell which bucket you’re in,” he says.
Along the way, Abraham picked up a bit of classic sales advice that he passes along to other folks doing the founder-led sales grind. “At first, I felt like every lead could not be dropped, but I’ve changed my perspective over time. The best answer you can get in sales is a yes, the second-best answer is no. It’s the maybes that will kill you.”
And remember, you’re never really done selling. “When selling to the Fortune 10, there were three different points where we got some form of verbal yes and felt like we had the deal secured and, in retrospect, there were still so many stages to go. A signed contract just leads to you needing to deliver at onboarding, and even after onboarding you need to keep nurturing the relationship so that they’ll renew,” Abraham says.
Founder-led sales: Relationships matter
Reducto only recently opened up a search for the first sales hire — after Abraham closed millions in ARR as a one-man GTM band. “I learn a lot from the sales calls I hop on, so I didn’t want to outsource that too early. It felt almost deadly for the company to lose that pulse with customers,” he says.
And so far, that lack of sales expertise hasn’t slowed Reducto down — if anything, it’s been a competitive edge. “I don’t think the reason why we’re selling at the scale we are is a matter of wordsmithing the right answer. It’s just an obsession with getting it right, and I think customers feel that. It’s hard to quantify, but caring really matters,” he says.
He points us back to that pivotal moment of signing the Fortune 10 customer quite early on. “We certainly didn’t have full feature parity compared to other vendors that they must have talked to. But they weren’t only betting on the immediate state of the product; they were betting on me and Raunak,” says Abraham. “When they brought up a missing feature or had a complaint about an edge case, we would have it fixed for them an hour later.”
Your energy as a founder is contagious. When people see how much you care about the product, they start to care about it more, too.
He imparts a bit of advice for other technical founders trying to sell to enterprise — think emotion, not logic. “You’re probably someone who approaches your own buying decisions from a very rational perspective — you choose whatever you think will perform the best. That’s how our startup customers treat Reducto, it’s very objective,” he says. “But with enterprise sales, it’s much more relationship-driven in a way that surprised me as someone who didn’t come from a sales background.”
Legal teams, security teams, procurement teams — there’s an enterprise gauntlet that startups must face down to get a contract signed. “The only way that you will navigate that efficiently is to have at least one person who’s willing to champion that process for you. Spending time with that champion and making them excited not just about the products, but about working with you is way more important than I would have known two years ago,” he says.
To put it into perspective, he borrows a piece of advice from former founder and First Round partner Liz Wessel (who led our seed investment in Reducto). “I remember her telling me that she built such deep relationships with her own early customers as a founder that she would feel comfortable inviting them to her wedding. The more I’ve thought about selling from that lens, the easier everything else becomes.”
And while he’s learned this lesson most painfully from founder-led sales, it’s one that applies more broadly: embrace the fumbling phase.
“On day one of starting this company, I had reticence about doing things that I wasn’t comfortable with because of the fear of doing it poorly. The first demo call would be daunting because I hadn’t done it. Prompting a user to pay feels weird,” Abraham says. “But over the course of building a company and sucking at something, you do it over and over again until you almost forget how bad you were at the start. As a founder, you need to realize the things you suck at are not a negative reflection on you. It’s an opportunity to unlock a new lever for the company.”
It’s very important to learn how little it matters to fail on the micro level.
The Benchmark: Show, don’t tell
As a team, Reducto was obsessive about being best-in-class. “We were always very methodical about validating, benchmarking and testing what we were doing,” says Abraham.
But in a sea of “We’re the most accurate parsing tool!” claims, how do you stand apart to customers? Reducto kept butting up against this exact problem.
“Our space isn’t differentiated by feature lists — accuracy differences for our customers directly translate to fewer mistakes that would otherwise impact everything downstream. But we kept finding on first sales calls the buyer would often say, ‘I’ve taken so many calls this week, what’s actually different about you guys?’” he says. “Once they spent a lot of time looking at our outputs against others, they would conclude that Reducto worked better.”
But in startups, time is not always on your side — Reducto needed to find a way to close the education gap and for customers to get to that lightbulb moment faster. “We wanted to remove all the burden on customers to have their eng team come up with data points, the scoring framework, everything they needed to evaluate different tools.”
Building a comprehensive benchmark would be no weekend project (in total, it took about three weeks), but Abraham had firm conviction that it was worth the undertaking. Reducto employed a team of PhD-level human labelers who manually annotated 1000 complex table images from a diverse set of publicly available documents (a mix of examples with different structures, text density, and language). Next, they came up with a fair scoring system (more on that here) and then open-sourced the benchmark.
People with zero data points will still tell you that their models are state of the art. We wanted to show our work — here’s how we back up what we’re claiming.
It’s rare that a go-to-market move flips the switch so quickly — it’s usually more of a slow burn. But Reducto saw this tactic bear fruit almost right away. “We had people reach out who hadn’t engaged with Reducto before because they saw the benchmark and appreciated the work that went into it. And we’ve also had people in the sales cycle specifically call out that the benchmark did the heavy lifting for them,” says Abraham.
While benchmarking is quite common in ML research organizations, Reducto was the first in the parsing space to open-source their benchmark. “Now it’s becoming trendy and I see a lot of other startups releasing their own versions. I don’t mind, more data is always good,” Abraham says. “Reducto is solving an incredibly hard problem. We constantly measure ourselves so we can be even better.”
Growing fast, hiring slow
Reducto has intentionally kept the team extremely small — in fact, they crossed $1M in ARR when they were just a team of four (and they’re still around a dozen today). “We care a lot about the efficiency of each person. One of our first hires was an ML researcher who did his entire PhD in document processing. Nailing that one hire was way more important to us than trying to assemble a 10-20 person engineering team,” says Abraham.
Part of this is preference (“having a large, unwieldy organization was never attractive to me and Raunak,” he says). But much like Abraham’s shrunken GTM team of one, he sees it as an upper hand. “If you’re an engineer in a 40-person organization focused on PDF processing, you probably have a very siloed sense of what your task is for that quarter for that year. But if you have a small set of people working on a very important problem, each person is forced to go deep,” he says.
It also means that each person is exceptionally close to customers and their urgency.
Every single decision and feature that we’re working on, we know there’s a customer waiting and counting on us to do it. The problem isn’t abstracted away into numbers on a dashboard.
Every startup taking on big established players must do everything they can to capitalize on this sense of urgency, Abraham says. “One of our large enterprise customers actually had an internal document processing team and had engineers staffed on the same problem. But at the end of the day, they ended up choosing Reducto because they saw that Reducto was getting better day over day — not month over month or year over year. That matters in ways you can’t quite put a label on.”
When deciding which legos to give away, Abraham abides by a simple mantra: “What’s the step in the pipeline that I’m not learning from anymore?”
The roadmap
While Reducto briefly toyed with the benefits of a more verticalized go-to-market strategy, they stuck with a broader appeal. “It would have been easier to be a healthcare-specific doc processing company. But being able to handle different industries has made the product better. We value seeing data points that we’ve never had to deal with, and that’s only possible if you’re building a horizontal product,” Abraham explains.
That means the array of features Reducto could build is quite vast (and perhaps a bit daunting). Here’s how Abraham delineates what gets added to the roadmap: “There are features that are fundamentally transferable and features that are point solutions. If someone comes to us from healthcare and says X type of document needs to work, the limitation to doing that might be our table parsing. Improving our table-parsing is not a healthcare-specific issue, it will improve our product for customers in finance and insurance, too,” he says. “But if a customer is looking for a niche file extension and that’s the only thing they care about getting parsed, given our team size, that’s not the customer we’ve decided to chase. We’re very upfront when we’re not the right solution for them.”
He shares a specific example: “For months, we would have people reach out and say, ‘I have equations as part of these research papers, can Reducto parse this?’ And our answer would be no because we couldn’t prioritize building it at the quality bar we’ve set for ourselves, we needed to fix fundamental issues like table parsing because the dollar value is orders of magnitude higher.”
We mentioned up top that Reducto has a Slack channel with just about every customer — here’s where that comes in handy, too. “We can see where dozens of customers are asking for this one thing and it’s meaningfully impeding the value that they expect from us,” says Abraham.
And while parsing was the initial wedge that drove hordes of folks to sign up for Reducto, it’s now table stakes. “RAG as a paradigm introduced this need to think about chunking. It wasn’t just, ‘Can you parse my documents?’ It was, ‘After you parse it, what do I feed into my VectorDB?”
But today, Abraham says he almost never gets asked about chunking. “We’ve built this entire pipeline for chunking that’s incredible, but we’ve come to realize that what our users want to do is beyond just reading the document. So today we have features that go beyond that, like process automation on top of Reducto. People will classify their documents, split them, and they’ll do structured extraction on those documents, all within Reducto.”
As he thinks about divvying up the work ahead, parsing still gets the biggest slice of the pie. “Every single one of those things is downstream of our ability to parse them really well. Even though we have these other features, a majority of our time is spent thinking about that core,” he says.
The path forward
Back when Abraham and Chowdhuri started the company, they made an initial promise to stick with it for two years. At the time, it seemed like a threshold where if they had stopped before that mile-marker, they probably wouldn’t have tried hard enough, Abraham says.
“If you asked us two years ago if we would be excited to work on PDF processing, the gut reaction would probably be no. But as we’ve gone deeper, we’ve found a love for what we’re doing. The moments of pride where our customers put us head to head with competitors and say that our app is 10x faster — that’s enough for us to want to work on this for who knows how long,” Abraham says.
If you’re a Reducto customer, you can text or Slack founders Adit Abraham and Raunak Chowdhuri at any time.
It’s a simple, but effective instrument for keeping a pulse on customers, as close to sidling up to their desk as an early-stage startup can get. Abraham credits it as one of the most powerful tools in the young startup’s toolbelt. But it’s, admittedly, not the sort of splashy advice you typically look for when trying to parse a breakthrough company’s path to product-market fit — and that’s precisely the point.
“There’s a disparity when founding a company. We hear inspirational stories of these grand visions and the big-picture ideas. So when you’re a new founder and you’re thinking about how to talk to customers, it doesn’t feel instinctively right to just reach out to people on Slack. It feels like you should be doing something flashier or cooler,” he says. “But in the beginning, it only matters whether or not people are using your products.”
He likens novice founders to newbie exercise enthusiasts — they both tend to overcomplicate things:
If you ask a bodybuilder for advice to get fit, they would tell you very simple things: Focus on your nutrition, go to the gym, prioritize sleep. But you see folks trying to get into fitness will instead agonize over figuring out the perfect training regimen and complex routines.
Most company profiles chronicle the rise to success long after it’s already been cemented. While we’ve shared a number of stories in our “Paths to PMF” series featuring companies that have long since found extreme product-market fit, we think there’s equal value in mining the stories of startups that are a few clicks earlier in their PMF journeys. As time removes you further away from the early days, recollections get hazy, and tactical steps slowly morph into broad platitudes like “stay close to your customers.”
But before we dive into the nitty-gritty of how they made it all happen, a brief explainer of the problem that Reducto is solving: 80% of the world’s business data is in unstructured file formats (like PDFs or Excel spreadsheets). If you’re trying to build a reliable AI application for handling, say, insurance claims, health records or financial statements, you’ll find that even the best LLMs start to hallucinate trying to read these unstructured documents. Multi-column layouts get jumbled together, key figures are ignored, and tables are a nightmare. Reducto reads these complex docs the way humans do and creates LLM-ready inputs with remarkable accuracy.
And while Reducto is still early, the trajectory has been remarkable, and as seed investors First Round has been ringside as Abraham and co-founder Chowdhuri hurtle past milestones. In just one year, they went from a pre-seed to a seed round to their just-announced $24.5M Series A and nabbed a Fortune 10 enterprise customer. If you printed out every page that Reducto has parsed so far, it would be something like 3.5x the size of Mount Everest.
In this exclusive interview, we pluck out some of Abraham’s biggest lessons on building Reducto so far, while they’re still incredibly fresh. And while there’s advice that follows specifically for founders building in AI, other early startup folks will have plenty to jot down in their notebooks — from Abraham’s one-man GTM show, to how they’re keeping the team quite lean. Let’s dive in.
Early brushes with entrepreneurship
Abraham’s entrepreneurial ambitions can be traced back to perhaps an unlikely place — the mobile game Flappy Bird. “Games were the thing that drew me to computers and Flappy Bird was what made me interested in programming,” he says. So as young, ambitious high schoolers, Abraham and his friend started making their own apps. The rumors that Flappy Bird’s creator was netting $50,000 a day off the app didn’t hurt, Abraham laughs. “We had these moments of like, ‘Forget school, we should just be making apps and we won’t even need to go to college.’”
So the duo spun up a goal-setting app, and while it didn’t make Abraham a newly minted app mogul, it did launch something of a side hustle. “I learned more about marketing than I did about building apps. We started making Instagram pages to promote the app, where we would post summaries of books like ‘How to Win Friends and Influence People’ and those pages started taking off. Even when the company shut down, I started running Instagram pages for other companies,” he says.
And while the goal-setting app might have fizzled out, Abraham’s entrepreneurial spark was just starting to ignite. He went on to M.I.T., where he eventually joined the Accelerator program and started building Sidewalk, a Shopify plugin that would recommend products from other stores during the checkout process.
But it was a chance classroom encounter that would truly turn the tides for Abraham’s founder path.
The team up
Abraham, then a junior at M.I.T., had walked into the first day of a graduate-level ML course with a heavy dose of undergrad imposter syndrome. So imagine his surprise when a freshman was standing at the front of the lecture hall. “The professor introduced Raunak to the class and told us he would be walking us through the first problem set. It was kind of obscene to be sitting there and see this person who had been doing ML research since he was 12,” Abraham recalls (Chowdhuri actually had 100 citations to his name before he even got to campus).
The two struck up a friendship, joined the same living group at M.I.T., and stayed in touch after Abraham graduated. After running Sidewalk for a year, he joined the ads team at YouTube. And frankly, he was bored. “We were both in a transition phase. Raunak had wrapped up his previous company and was going to be graduating from M.I.T. soon. I was vocal about being ready for something new,” Abraham says. So they did what plenty of other young engineers do — they started applying to hackathons together (they actually won Anthropic’s hackathon fresh off the release of Claude).
After a few of these hackathons (and a few trophies) under their belts, they finally broached the topic: Did they want to start a company together? “At least for me, it was an immediate yes. There was no question in my mind that Raunak would be an exceptional person to work with,” Abraham says. Their areas of interest also tipped the scales — Abraham was more drawn to the GTM and product side of the business, Chowdhuri was deeply technical.
At the time, I wasn’t focused on the idea that we would work on or whether we could get funding. I was just excited to build with this person.
So they plucked an idea from a past hackathon project Chowdhuri had worked on, and applied to YC, getting accepted into the Winter ‘24 batch.
The initial idea
Their initial YC application was far removed from PDF processing — it was a long-term memory solution for language models. And the co-founding duo gave themselves a simple rule: “We decided if we got into YC, we would at least stick with the idea for the duration of the batch. We were worried about just pivoting around aimlessly,” Abraham says.
But there were some early signals that, while attention-grabbing, their idea didn’t have legs. “It did go somewhat viral on X — people thought it was cool to see a language model bring up something you mentioned in the past,” he says.
Hundreds of folks asked to be onboarded to the product, then called Remembrall (“Harry Potter” fanatics will notice a pattern with these company names), but early onboarding calls hinted that they were a bit too early to the space.
They had conversations with enthusiasts who would maybe try adding it to their AI application. “It was never the case where people were saying, ‘I’ve noticed that customers get angry at my product because of X, Y, and Z and Remembrall can fix it.’ That was the first signal. The second signal was that people were maybe willing to pay $10-20 a month because they found the idea interesting. But no one’s product team needed it,” he says.
We set up demo calls with everyone who expressed interest and we asked what their use case was and what problems the product would solve for them. And we would very rarely get a real answer. It was more of a ‘This seems cool’ curiosity.
But buried within these somewhat directionless enthusiast chats was an ember of an idea that needed fanning.
The pivot
“One of the most common feature requests for Remembrall was, ‘Hey, you’re managing my user’s chat history, can you manage the files that they’re uploading as well?” Abraham says. “We went into our projects expecting to help teams with fun ML problems, but very quickly learned that one of the biggest bottlenecks across most pipelines was actually well before retrieval or generation.”
So the co-founders built a very simple solution — they would upload the docs, parse them using an external parsing solution, and then chunk the information for you. “It’s embarrassing to look back on it now, it was a very ugly Streamlit app with a super simple document segmentation tool. It would do nothing other than split your documents into little boxes of content. I can’t stress this enough — it was literally a weekend project that we threw together,” says Abraham.
But they posted it on YC’s forum anyway, and the response was immediate and overwhelming. “We started getting replies, ‘This is better than what I’m getting from Textract. Is this a hosted API? Where’s the Stripe link?’ The pull was much stronger than everything else we had worked on,” he says.
“We knew how annoying it was to build a PDF processing pipeline. We knew how much time we were putting into post-processing that content. So when other people mentioned the same painful experience, it was a clear aha moment,” he says.
Unlike the “seems cool!” enthusiasm for extended LLM memory, there was a distinct need and a massive pain point. “These companies are not trying to spend many hours of engineering time on PDF processing, but it’s the bottleneck that’s stopping them from building the things that actually matter. If we can be that ingestion team for our customers and take that problem off their plate, that’s clearly very valuable,” he says.
We had spent enough time exploring things that were not resonating. In comparison, it felt like we were getting punched in the face by this new idea — it was something people cared about an order of magnitude more.
The competition conundrum
But let’s be clear here, while the team was getting inundated with signals that they had sniffed out a winning idea, they were somewhat hesitant to wholly embrace this new direction. For one thing, neither co-founder had expected to build a PDF processing company. And there was an initial gut fear that they wouldn’t be able to build a moat quickly enough.
“PDF processing is not a new space,” says Abraham. (Brace yourselves here — PDFs have been around longer than both Reducto founders have been alive.) “Every day we were looking into this idea, we’d find a new, different solution that claims to solve the problem. But the way we thought about it was that everything on the market will get you part of the way there. Nothing we tried was at the accuracy level for what we wanted, and clearly not at the level our customers wanted.”
Here’s how he explains the difference in layman’s terms: “There was an old era of parsing PDFs using the metadata and just trying to read the file itself. Our insight was that these documents were made for humans, not machines, to read. Every little visual cue, like a gap between two paragraphs, is me telling you, ‘Hey, this is a new semantic piece of information.’ Or a tab structure in a list tells you, ‘This is a sub-idea of that parent idea,’” Abraham says. “We wanted to read these documents like a human would. And that is a very deep problem, but thinking about it from a first principles lens unlocks a whole range of different ways to improve parsing.”
I think younger founders like ourselves are more afraid of competition than they should be. Our question wasn’t, ‘Is the market big enough?’ The question was, ‘Can we be better in a meaningful way?’
Once the founders felt confident that the answer was yes, they put blinders on. “We didn’t want to spend a year or two of our lives parading around pretending to build a company. We just wanted to focus on this one thing,” Abraham says.
The MVP
Today, Reducto has six different vision models to ensure accurate outputs with any document, but they started with just one. “We trained the first model for layouts, breaking down a document into sub-regions. As a human, when you look at a multi-column layout, you know how to read it. But that’s not immediately obvious if you have text just sitting there. You’ll read the text from left to right, you might merge the columns together, etc. And so we tried to break it down into how we understand the paragraph structure and the position,” he says. “We actually see PDFs as just images, and so we convert them to images. It’s not about PDF as a standard, it’s about documents and human content overall.”
They cobbled together an initial pipeline and paid UI barely any thought at all. They managed to officially launch just two weeks after that initial post on YC’s internal forum — although they were hesitant to do so, despite prodding from folks to get the tool out there.
We got a lot of advice to launch early and not be afraid of failure. But we knew we were in a space where if we launched too early and the product wasn’t good, it would just be a waste of everyone’s time.
“People already had PDF processing solutions, Reducto needed to be decidedly better. Then once we cleared that bar, we would try to launch as soon as possible,” says Abraham.
They kept the launch simple. “We just put up a post on our social channels saying essentially, ‘Hey, we know PDFs suck. Reducto’s building vision models to address that. Try it out for yourself,” says Abraham.
But here’s the genius behind their launch buzz: “We had a playground with a pre-populated document, but you could also upload your own file. People usually had some sort of file that they’d seen fail with parsing all the time. Seeing it work with Reducto immediately sparked attention.”
Inbound interest poured in from startups, consultancies, even BigCos (by the time demo day rolled around in April, Reducto was getting 55K in monthly traffic). But Abraham and Chowdhuri were intentionally choosy about who to initially partner with.
The first customers
To sidestep the enthusiast problem that they had faced with Remembrall, the co-founders wanted to focus their attention on an ICP that felt the pain most acutely. That meant industries where accuracy was imperative — like finance, legal, or insurance. Making an error on someone’s insurance claim isn’t just a silly typo.
And while they started having some initial conversations with enterprises that had expressed interest, they mainly targeted startups. “They would be able to integrate us by the end of the week,” Abraham explains.
The onboarding experience was white-glove, and even today, Reducto does not have fully self-service onboarding. “We still manually onboard every customer because we learn a lot from the process,” he says.
It was also a quality control mechanism. “We didn’t want to be in a situation where someone integrates us into their production environment and all of a sudden our infra wasn’t ready to handle the next million pages we were hit with. We wanted a failsafe on the customer experience,” Abraham says.
Pricing for acute pain
To sketch out their pricing model, Abraham balked at the more expected usage-based approach and instead went with a tiered system that starts at $300/month. “Our pricing tiers assume that you’re processing at a certain volume. The startups that we work with tend to be processing above 15,000 pages a month, at minimum. In some cases, hundreds of thousands or even millions,” he says.
Part of this thinking stems from a bit of scar tissue from the initial enthusiast reaction to Remembrall — and the $10/month price tag that came with it. “We want this to be something that’s tangibly impactful and real infrastructure for companies to build AI products. In order to do that, we wanted use cases that matter to the end customer.”
Enterprise comes calling
While their preliminary focus was on getting quick-acting startups onboarded into the tool, it didn’t take long before enterprises took notice of Reducto — with one seemingly landing in their lap just a few days after publicly launching. “A Fortune 10 company actually tried out the playground and signed up for a demo call,” Abraham says.
After some exploratory calls, at one point during the sales process they brought 14 of their engineers to meet with the Reducto co-founders for an entire day — the signals of developing PMF can’t get much clearer than that.
“Their entire team just sat with us to learn more. The depth at which they cared and had thought about the details showed that it mattered to them. I didn’t feel that we had product-market fit until our enterprise conversations started getting escalated in a way we didn’t expect,” says Abraham. Eventually (after 154 days total, over 20 hours worth of meetings, an acquisition attempt and a few hundred emails), the Fortune 10 signed on.
But other enterprise deals were an even slower burn. Abraham’s former college roommate was on the team at Scale AI and connected them to Reducto, but it would take another year before they eventually signed the deal. Here’s Abraham’s advice for maintaining focus — and patience — during these prolonged sales cycles.
“The clearest signal is how much your champion seems to care. There are enterprise deals where I’m reaching out to the champion almost begging them to get on a call and they never respond to me. Other times, there are just hoops you have to jump through, maybe their legal team is taking forever, but your champion still seems very engaged. You can clearly tell which bucket you’re in,” he says.
Along the way, Abraham picked up a bit of classic sales advice that he passes along to other folks doing the founder-led sales grind. “At first, I felt like every lead could not be dropped, but I’ve changed my perspective over time. The best answer you can get in sales is a yes, the second-best answer is no. It’s the maybes that will kill you.”
And remember, you’re never really done selling. “When selling to the Fortune 10, there were three different points where we got some form of verbal yes and felt like we had the deal secured and, in retrospect, there were still so many stages to go. A signed contract just leads to you needing to deliver at onboarding, and even after onboarding you need to keep nurturing the relationship so that they’ll renew,” Abraham says.
Founder-led sales: Relationships matter
Reducto only recently opened up a search for the first sales hire — after Abraham closed millions in ARR as a one-man GTM band. “I learn a lot from the sales calls I hop on, so I didn’t want to outsource that too early. It felt almost deadly for the company to lose that pulse with customers,” he says.
And so far, that lack of sales expertise hasn’t slowed Reducto down — if anything, it’s been a competitive edge. “I don’t think the reason why we’re selling at the scale we are is a matter of wordsmithing the right answer. It’s just an obsession with getting it right, and I think customers feel that. It’s hard to quantify, but caring really matters,” he says.
He points us back to that pivotal moment of signing the Fortune 10 customer quite early on. “We certainly didn’t have full feature parity compared to other vendors that they must have talked to. But they weren’t only betting on the immediate state of the product; they were betting on me and Raunak,” says Abraham. “When they brought up a missing feature or had a complaint about an edge case, we would have it fixed for them an hour later.”
Your energy as a founder is contagious. When people see how much you care about the product, they start to care about it more, too.
He imparts a bit of advice for other technical founders trying to sell to enterprise — think emotion, not logic. “You’re probably someone who approaches your own buying decisions from a very rational perspective — you choose whatever you think will perform the best. That’s how our startup customers treat Reducto, it’s very objective,” he says. “But with enterprise sales, it’s much more relationship-driven in a way that surprised me as someone who didn’t come from a sales background.”
Legal teams, security teams, procurement teams — there’s an enterprise gauntlet that startups must face down to get a contract signed. “The only way that you will navigate that efficiently is to have at least one person who’s willing to champion that process for you. Spending time with that champion and making them excited not just about the products, but about working with you is way more important than I would have known two years ago,” he says.
To put it into perspective, he borrows a piece of advice from former founder and First Round partner Liz Wessel (who led our seed investment in Reducto). “I remember her telling me that she built such deep relationships with her own early customers as a founder that she would feel comfortable inviting them to her wedding. The more I’ve thought about selling from that lens, the easier everything else becomes.”
And while he’s learned this lesson most painfully from founder-led sales, it’s one that applies more broadly: embrace the fumbling phase.
“On day one of starting this company, I had reticence about doing things that I wasn’t comfortable with because of the fear of doing it poorly. The first demo call would be daunting because I hadn’t done it. Prompting a user to pay feels weird,” Abraham says. “But over the course of building a company and sucking at something, you do it over and over again until you almost forget how bad you were at the start. As a founder, you need to realize the things you suck at are not a negative reflection on you. It’s an opportunity to unlock a new lever for the company.”
It’s very important to learn how little it matters to fail on the micro level.
The Benchmark: Show, don’t tell
As a team, Reducto was obsessive about being best-in-class. “We were always very methodical about validating, benchmarking and testing what we were doing,” says Abraham.
But in a sea of “We’re the most accurate parsing tool!” claims, how do you stand apart to customers? Reducto kept butting up against this exact problem.
“Our space isn’t differentiated by feature lists — accuracy differences for our customers directly translate to fewer mistakes that would otherwise impact everything downstream. But we kept finding on first sales calls the buyer would often say, ‘I’ve taken so many calls this week, what’s actually different about you guys?’” he says. “Once they spent a lot of time looking at our outputs against others, they would conclude that Reducto worked better.”
But in startups, time is not always on your side — Reducto needed to find a way to close the education gap and for customers to get to that lightbulb moment faster. “We wanted to remove all the burden on customers to have their eng team come up with data points, the scoring framework, everything they needed to evaluate different tools.”
Building a comprehensive benchmark would be no weekend project (in total, it took about three weeks), but Abraham had firm conviction that it was worth the undertaking. Reducto employed a team of PhD-level human labelers who manually annotated 1000 complex table images from a diverse set of publicly available documents (a mix of examples with different structures, text density, and language). Next, they came up with a fair scoring system (more on that here) and then open-sourced the benchmark.
People with zero data points will still tell you that their models are state of the art. We wanted to show our work — here’s how we back up what we’re claiming.
It’s rare that a go-to-market move flips the switch so quickly — it’s usually more of a slow burn. But Reducto saw this tactic bear fruit almost right away. “We had people reach out who hadn’t engaged with Reducto before because they saw the benchmark and appreciated the work that went into it. And we’ve also had people in the sales cycle specifically call out that the benchmark did the heavy lifting for them,” says Abraham.
While benchmarking is quite common in ML research organizations, Reducto was the first in the parsing space to open-source their benchmark. “Now it’s becoming trendy and I see a lot of other startups releasing their own versions. I don’t mind, more data is always good,” Abraham says. “Reducto is solving an incredibly hard problem. We constantly measure ourselves so we can be even better.”
Growing fast, hiring slow
Reducto has intentionally kept the team extremely small — in fact, they crossed $1M in ARR when they were just a team of four (and they’re still around a dozen today). “We care a lot about the efficiency of each person. One of our first hires was an ML researcher who did his entire PhD in document processing. Nailing that one hire was way more important to us than trying to assemble a 10-20 person engineering team,” says Abraham.
Part of this is preference (“having a large, unwieldy organization was never attractive to me and Raunak,” he says). But much like Abraham’s shrunken GTM team of one, he sees it as an upper hand. “If you’re an engineer in a 40-person organization focused on PDF processing, you probably have a very siloed sense of what your task is for that quarter for that year. But if you have a small set of people working on a very important problem, each person is forced to go deep,” he says.
It also means that each person is exceptionally close to customers and their urgency.
Every single decision and feature that we’re working on, we know there’s a customer waiting and counting on us to do it. The problem isn’t abstracted away into numbers on a dashboard.
Every startup taking on big established players must do everything they can to capitalize on this sense of urgency, Abraham says. “One of our large enterprise customers actually had an internal document processing team and had engineers staffed on the same problem. But at the end of the day, they ended up choosing Reducto because they saw that Reducto was getting better day over day — not month over month or year over year. That matters in ways you can’t quite put a label on.”
When deciding which legos to give away, Abraham abides by a simple mantra: “What’s the step in the pipeline that I’m not learning from anymore?”
The roadmap
While Reducto briefly toyed with the benefits of a more verticalized go-to-market strategy, they stuck with a broader appeal. “It would have been easier to be a healthcare-specific doc processing company. But being able to handle different industries has made the product better. We value seeing data points that we’ve never had to deal with, and that’s only possible if you’re building a horizontal product,” Abraham explains.
That means the array of features Reducto could build is quite vast (and perhaps a bit daunting). Here’s how Abraham delineates what gets added to the roadmap: “There are features that are fundamentally transferable and features that are point solutions. If someone comes to us from healthcare and says X type of document needs to work, the limitation to doing that might be our table parsing. Improving our table-parsing is not a healthcare-specific issue, it will improve our product for customers in finance and insurance, too,” he says. “But if a customer is looking for a niche file extension and that’s the only thing they care about getting parsed, given our team size, that’s not the customer we’ve decided to chase. We’re very upfront when we’re not the right solution for them.”
He shares a specific example: “For months, we would have people reach out and say, ‘I have equations as part of these research papers, can Reducto parse this?’ And our answer would be no because we couldn’t prioritize building it at the quality bar we’ve set for ourselves, we needed to fix fundamental issues like table parsing because the dollar value is orders of magnitude higher.”
We mentioned up top that Reducto has a Slack channel with just about every customer — here’s where that comes in handy, too. “We can see where dozens of customers are asking for this one thing and it’s meaningfully impeding the value that they expect from us,” says Abraham.
And while parsing was the initial wedge that drove hordes of folks to sign up for Reducto, it’s now table stakes. “RAG as a paradigm introduced this need to think about chunking. It wasn’t just, ‘Can you parse my documents?’ It was, ‘After you parse it, what do I feed into my VectorDB?”
But today, Abraham says he almost never gets asked about chunking. “We’ve built this entire pipeline for chunking that’s incredible, but we’ve come to realize that what our users want to do is beyond just reading the document. So today we have features that go beyond that, like process automation on top of Reducto. People will classify their documents, split them, and they’ll do structured extraction on those documents, all within Reducto.”
As he thinks about divvying up the work ahead, parsing still gets the biggest slice of the pie. “Every single one of those things is downstream of our ability to parse them really well. Even though we have these other features, a majority of our time is spent thinking about that core,” he says.
The path forward
Back when Abraham and Chowdhuri started the company, they made an initial promise to stick with it for two years. At the time, it seemed like a threshold where if they had stopped before that mile-marker, they probably wouldn’t have tried hard enough, Abraham says.
“If you asked us two years ago if we would be excited to work on PDF processing, the gut reaction would probably be no. But as we’ve gone deeper, we’ve found a love for what we’re doing. The moments of pride where our customers put us head to head with competitors and say that our app is 10x faster — that’s enough for us to want to work on this for who knows how long,” Abraham says.
