Adit Abraham Found a Business in the Documents Everyone Else Hated

The first version of Adit Abraham's entrepreneurial life had a cartoon bird in it. As a high school student, he read that the creator of Flappy Bird was earning $50,000 a day. Abraham and his best friend reached the conclusion teenagers are uniquely equipped to reach: forget the normal path, make apps. They built a goal-setting app. It did not produce Flappy Bird money, but it left him with a more durable asset. Trying to distribute it taught him that building and selling were parts of the same job.

The pair created Instagram pages that posted compact lessons from books such as How to Win Friends and Influence People. The pages took off more readily than the app. Even after the product disappeared, Abraham ran pages for other companies. His first startup lesson arrived early and without ceremony: attention and need are different things, and sometimes the marketing experiment outlives the thing being marketed.

At MIT, he studied computer science and engineering, worked on machine-learning research at the Media Lab and built Sidewalk, a Shopify product-recommendation project, through an accelerator. He also walked into a graduate machine-learning class as a junior and found a freshman preparing to teach the first problem set. That freshman was Raunak Chowdhuri, who had already been doing ML research for years.

Choose the person before the pitch

Abraham remembers feeling the ordinary undergrad version of imposter syndrome. Chowdhuri, standing at the front of the room, made an extraordinary first impression. They became friends, joined the same MIT living group and kept in touch after Abraham graduated. Abraham ran Sidewalk for about a year, then joined Google as a product manager. He worked first on YouTube ads, where releases met concrete business targets, and later on Search, where the path between his work and the result felt less direct.

He has since resisted turning that contrast into a slogan about startups defeating big companies. His own conclusion is narrower and more useful: decide what kind of work helps you learn. A role with clear responsibility inside a large company can be excellent. A role that leaves you disconnected from the outcome can be a reason to move. He left because a problem and a collaborator gave him energy, and because not trying would have become its own regret.

“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.”Adit Abraham, on starting with Raunak Chowdhuri

The two tested the relationship through hackathons. Fresh off Anthropic's release of Claude, they won one together. When the conversation finally turned to starting a company, Abraham's answer was immediate. The idea mattered less than the evidence they had already collected about each other: they liked working together, their instincts were complementary, and Abraham admired Chowdhuri's technical depth.

A popular idea with no urgency

Their Y Combinator application proposed Remembrall, a long-term memory layer for language models. The demo spread on X. Hundreds of people asked to be onboarded. On the surface, this looked like the kind of response founders are told to chase. The calls sounded different. Prospective users found it interesting and might pay $10 or $20 a month, but few could name a product problem that demanded a solution now.

Abraham and Chowdhuri had promised themselves they would not drift through ideas during Y Combinator's Winter 2024 batch. Still, one repeated feature request carried more weight than the praise. If Remembrall could manage chat history, customers asked, could it also manage the files users uploaded? The bottleneck sat earlier in the AI pipeline. Before a model could retrieve or reason over information, someone had to make a PDF, scan or spreadsheet legible to the system.

They assembled an initial pipeline around a single model that identified a page's regions and reading order. The insight was visual and almost embarrassingly human: a gap between paragraphs, an indented list and two columns all carry meaning. Software that extracts text without seeing those relationships can return every word and still lose the document.

The first paid interface received little aesthetic attention. The first Reducto website was essentially a form. Yet the product had a more persuasive feature: a playground where visitors could upload the documents that had defeated other parsers. Abraham did not need to insist on accuracy when the buyer could supply the test. A few days after the public launch, a Fortune 10 company booked a demo.

“Our question wasn't, ‘Is the market big enough?’ The question was, ‘Can we be better in a meaningful way?’”Adit Abraham

The contract inside the calendar

The inbound meeting was the beginning, not the breakthrough. At one point, 14 engineers from the prospective customer spent a day with the two founders, examining what Reducto could and could not do. The process lasted 154 days, consumed more than 20 hours of meetings and hundreds of emails, and included several moments that sounded like yes before a contract actually existed.

Abraham learned to distinguish a delayed deal from a dead one by watching the champion. Legal and procurement can move slowly while the person with the problem remains engaged. Silence from the supposed champion says something else. His compact version of the lesson is that yes is the best answer, no is second best, and maybe is the answer that consumes the calendar.

He also stayed on the calls because selling exposed product truth. Reducto manually onboarded every early customer. Abraham configured subscriptions himself. When an outside labeling team could not reach the required accuracy, he drew boxes around document regions. None of this looked like the work of a chief executive in a funding announcement. Each task shortened the distance between a customer's complaint and a product decision.

Proof before adjectives

Document AI is a market where every vendor can claim high accuracy. Abraham's response was to make evaluation part of distribution. Reducto assembled and open-sourced a table benchmark built from 1,000 complex images annotated by expert labelers. The company showed its scoring method and invited comparison. The benchmark did the explanatory work that a feature list could not.

The same logic shaped the organization. Reducto crossed $1 million in annual recurring revenue with four people and closed millions before adding a formal sales team. One early ML hire had completed a doctorate focused on document processing. Abraham and Chowdhuri preferred a small group in which an engineer could see the customer waiting for a feature, rather than a larger team where the request arrived as a number on a dashboard.

First Round led an $8.4 million seed round announced in October 2024. Benchmark led a $24.5 million Series A the following April. Six months later, Reducto announced a $75 million Series B led by Andreessen Horowitz, bringing disclosed funding to $108 million. By then, the company said it had processed more than a billion pages, and monthly volume had increased sixfold since the Series A.

The product had also widened beyond its first parser. Reducto added tools to split, extract and edit documents, plus a Studio interface for building and evaluating full pipelines. The expansion followed a sequence Abraham favors: solve one broad, transferable problem well, then take on the next. Layout understanding could help across industries. Table detection could do the same. Each capability widened the surface without replacing the standard underneath it.

The pile nobody photographs

Funding compresses time. Three rounds appear as three dates, and the chart rises cleanly. Abraham's more revealing stories are stubbornly physical: the unlabeled document on a screen, the first uncomfortable payment prompt, the customer call he was tempted to outsource, the on-premise deployment the small team had never attempted. They are moments in which competence did not arrive before the work. It arrived through repetition.

He tells founders to become less afraid of failing at the micro level. The first demo can be poor. The first sales conversation can feel unnatural. The relevant question is whether doing it again teaches the company something. Reducto's public success rests on a private tolerance for being temporarily bad at the next necessary job.

There is playfulness around the edges. The company and its predecessor both borrowed names from Harry Potter. Abraham's Y Combinator biography confesses to an unreasonable amount of Pokémon Showdown. On X, his bio reduces the CEO job to “document processor.” The joke works because it points back to the work. Strip away the category language, and Reducto helps a machine understand a file that a person can already see.

Abraham and Chowdhuri once promised each other they would give the company two years, enough time to know they had tried. Neither expected to develop affection for PDF processing. Depth changed that. A customer choosing their output in a head-to-head test became its own reward. The unlovable document turned out to contain a long runway, provided they kept reading it closely.