Document AI Must Catch What It Silently Leaves Out

A PDF was built to preserve a page for printing, not to help an AI agent understand what is on it. That mismatch is the problem Reducto is trying to solve, according to co-founder and CEO ? adit-abraham. Documents are a common way for organizations to store and exchange information, but the visual layouts that make them legible to people can be difficult for software to interpret reliably.

Abraham described the work as a deeper version of a problem that has been around for decades. He once met someone who had worked on PDF processing longer than Abraham had been alive, including on the original printer drivers for PDFs. But the needs have changed: a document is no longer only something a computer must reproduce faithfully on paper. Models increasingly need to read it well enough to do work that people used to do themselves.

That requires more than recognizing characters. People make rapid judgments about where one block of meaning ends and another begins, how a table connects information across rows and columns, and which visual details change the meaning of a page. Reducto’s target, Abraham said, is human-level fidelity: a system that can read documents the way a person would.

He compared the ambition to the gap between lane-assist features and a car that can drive autonomously from end to end. Existing tools may help with parts of document processing; the goal is to make the whole task dependable.

The company began with a different problem. Abraham said he and co-founder Raunak Chowdhuri entered Y Combinator with Remembrall, a long-term memory proxy for language models. On the day they got in, they told their YC contact they did not think Remembrall was what they would ultimately pursue. They had only a broad direction—“magical tools for developers”—and chose the Reducto name from a poll of possible names and available domains. The early memory work grew out of models that seemed powerful but had small context windows. Reducto’s focus shifted as the underlying problem and product opportunity changed.

A Hacker News post shown during the interview identified Remembrall as a “Long-term memory proxy for LLMs.” The Reducto website shown alongside it described the company as “The complete agentic document platform.” The juxtaposition captures the shift Abraham described: from helping models retain information to enabling them to work with documents.

Today, Abraham describes the company’s scope in terms of any interaction a model has with a document or other unstructured file. That includes reading and classifying information, reconciling it against other documents, and editing documents to produce an intended result. He is less interested in expanding into file storage for its own sake. The test for a new product area is whether Reducto can be the best in the world at the task—and whether taking it on improves the customer experience.

Document extraction has a distinctive failure mode: the system can omit information without making the omission obvious. In describing Reducto’s work on Deep Extract, Abraham said the team found a tradeoff between precision and recall. In his account, some specialist document-processing systems performed worse than frontier models when those models were given uncapped reasoning time. But the frontier models’ failures were often failures of recall: silently dropping rows or missing context.

That matters because a user may not know what was never extracted. An answer can look plausible while leaving out information from the source document. Abraham’s argument is not that a specialized harness should always replace a general-purpose model. Frontier models can be good choices for some tasks. But for a defined task, a specialized system may be more efficient, particularly when a harness can audit results and help catch omissions.

Reducto’s model strategy follows a related logic. Abraham said the company began with smaller traditional computer-vision models for routine processing and larger models for harder cases. The difficulty is that document workloads have a long tail: a page can look like ordinary text until an unusually complex table appears, and it may not be clear in advance which page will require more effort. At the time of the interview, Reducto was preparing to announce R1 “on Monday,” an in-house model intended to handle the task end to end—from capturing layout to making last-mile corrections. That description referred to the planned announcement at the time of the interview, not a claim about the model’s present availability.

Abraham framed the decision to build models in-house as a question of results, not identity. Customers care about performance, he said, not whether Reducto uses its own model or another provider’s. The practical target is the tradeoff among accuracy, cost, and latency. For specific tasks, Abraham said, post-trained models have repeatedly done better on all three dimensions, pushing Reducto’s stack toward in-house models. Broad, general-purpose foundation models still make sense for some tasks, though he described those uses as limited in scope at the time.

R1 was presented in the interview as a change in both performance and economics. Abraham said it would provide better accuracy than Reducto’s most accurate models at a per-page cost below that of its base models. It would also remove pricing multipliers customers had faced when turning on agentic features. He described the planned price as multiple times cheaper than the company’s previous pricing. These were launch claims made at the time of the interview; the source does not establish what pricing or availability is current. The objective, in his telling, was to improve the balance of accuracy, cost, and latency so more use cases could adopt the product in production.

Abraham also described an improvement to structured extraction that he said had been announced “today.” He said Reducto had improved citation performance and adherence to requested schemas as well as accuracy, while reducing cost. He put the previous all-in cost at more than four and a half cents per page and said the announced rate was half that, with performance also improving. He said the company’s North Star remained the most accurate document-processing platform, and that better cost efficiency followed from improving model performance rather than serving as a separate goal.

The proposed end-to-end model does not eliminate the need to monitor its output. For high-consequence workflows, Abraham said, there may never be a clean line between perfect and imperfect performance: ambiguity can remain even when a system reads the page well. Traceability—knowing where an extracted value came from—helps a reviewer check the result. Confidence scores can help direct attention to cases that warrant review. Internally, the company measures not only extraction accuracy but how reliably it can alert a human when an output is not perfect.

Abraham is wary of benchmarks that function mainly as marketing. He said companies often claim to be number one on their own evaluations, quickly making results noisy and difficult to compare. Reducto worked with Micron on a benchmark, he said, so the company would not choose the documents, set the ground truth, or manipulate the results. He described the aim as establishing an independent reference point; he did not detail what the benchmark tested.

He was careful not to present even that benchmark as definitive. At best, it gives potential users a directional signal and a place to begin. The most important evaluation, he said, is for customers to try systems on their own documents. That qualification matters for document workloads, where an average can conceal the unusual page, format, or layout that determines whether a particular process works.

Reducto also evaluates outputs after deployment. Abraham said the machine-learning team invests in continuous evaluation to identify results that are not necessarily wrong but are still imperfect in production. The purpose is to find patterns in failure, including cases outside the system’s familiar distribution, rather than simply collect individual bad examples. A healthcare form might include handwriting, for instance, but the more specific challenge could be handwriting on a tilted page, with marks extending beyond grid lines. Knowing where the system struggles helps the team decide what to include in later training runs.

This evaluation work also informs Reducto’s choice to serve multiple industries rather than build for just one. Abraham said many document tasks depend on common visual judgments: recognizing a new semantic block when paragraphs are separated by a gap, for example, or understanding how a table relates two dimensions of information. Improvements made against messy faxed tables in healthcare can also help with finance documents.

The shared foundation does not mean every workflow has the same requirements. Abraham noted that legal users may care about strikethroughs and underlines in ways healthcare users may not. His distinction is between objective interpretation and domain-specific meaning: Reducto wants basic parsing to be accurate and horizontal, so a system can identify what is on a page without needing to be trained separately for every industry. When a question depends on an organization’s definition—what a particular financial metric means, for example—the behavior should be promptable, steerable, and able to improve over time.

Reading information is only part of the work people do with documents. They also write into spreadsheets, modify Word files, redline contracts, and fill out applications. Reducto’s stated ambition is for models to handle more of that work reliably, rather than stop at extraction and parsing. Abraham said white-collar work had not yet had an equivalent of the “Claude Code moment”: a widely recognizable shift in which models take over a substantial, practical workflow.

The editing problem can be quite different from finding the right words. Abraham used document localization as an example. Models may be good at translation, but applying a translation to an existing document can require reworking its layout. Translating English into a right-to-left language such as Arabic can change text flow and cause content to spill onto another page. A system must preserve or adapt the document’s structure, not merely return translated sentences.

Reducto had launched a form-editing endpoint roughly six months before the interview, Abraham said. He said the company had more releases coming that he expected would mark a substantial step forward. Some editing tasks resemble computer use, while others are closer to code generation. In his view, both depend on understanding visual structure well enough to make the edit correctly.

The company does not sharply divide its machine-learning work into separate teams for parsing and editing. People may develop expertise in particular areas, but work across different problems sits within the same broader team. The discussion highlighted a recent hire, Liang, who had drawn attention for open-source office agents and was excited by document editing. Finding a person whose interests fit a problem was described as important.

The wider obstacle to automating office work is a mix of model capability and organizational context. Abraham said model intelligence was increasingly less likely to be the biggest bottleneck, given progress in long-horizon reasoning. But enterprise information is fragmented across systems, and important rules often exist as local knowledge rather than written procedures. Someone at a financial institution might know that a mark in the bottom-left corner of a paper form means it came from Kentucky, and therefore should trigger a particular action. That rule may never have been documented.

For models to operate across such processes, tooling must help collect that context and make it available in usable knowledge bases. Reducto works with companies developing end-to-end products in specific fields—Basis in tax filing and Harvey and Leya in legal work. Abraham said those specialist teams can discover what it takes to deliver value in their own contexts, while Reducto focuses on document capabilities that serve as part of the larger system.

Reducto’s earlier focus on agent memory has receded, but the underlying pattern of retrieving information has not disappeared. Abraham described many products as searches for relevant information, followed by combining it and reasoning or acting on it. That remains true whether a product visibly offers a search box or an agent traverses a file system to find the right context.

The label “RAG” often brings to mind a specific implementation: split documents into chunks, embed those chunks, and use similarity search to choose material for a model’s context. Abraham said some users resist the term because they are not building a chatbot. But an agent finding relevant material across multiple files, reasoning over it, and synthesizing an answer still follows the broad retrieval-and-generation pattern. The technical details may differ; the task is to locate information, assemble context, and use it to produce an outcome.

That pattern was also behind Reducto’s original memory work. Abraham said the company’s early direction grew out of language models that were already impressive but constrained by small context windows. As context windows grew, chunking became less of a headline concern, but search remained useful: many tasks require finding particular information across a collection rather than feeding everything to a model at once.

Reducto’s product work is tied to a hiring principle Abraham described as “obsession for your craft.” The relevant quality is not desire in the abstract, but the rigor someone brings to everyday work. He cited Yifei, an early hire and machine-learning researcher, as someone who examines data closely—down to individual PDF pages and the boxes drawn around them. At one point, Reducto had more data labelers than full-time employees, yet Yifei still inspected the data himself.

That attention should extend beyond model development. Abraham emphasized that engineers need to care about customer support and the long tail of production cases, not just features that make for visible launches. The company tries to hire people who want their work to be excellent, because that impulse is difficult to impose from the outside. In interviews, Abraham likes asking what a candidate is world-class at, including outside work. People deeply invested in a craft or hobby may bring the same intensity to a problem they care about professionally.

The company’s use of AI tools has spread without a mandate that every engineer use a specific coding assistant. Abraham credited co-founder Raunak’s example—he was described as unusually fluent with language models—and said the tools became attractive as people saw what colleagues could do with them. Reducto’s sales team, for example, started using AI tools to create custom websites and proposals for prospects.

As the company has grown, the founder’s job has changed. Abraham said the number of problems does not shrink when a company is doing well; the problems change. In the early days, the work included individual deals and support tickets. As the team expanded, he had to get used to not being constantly available on Slack and to trust the company to operate without immediate attention. Time with major customers remained important, while people-related work and culture became a larger part of the role.

Planning has lengthened somewhat without becoming a large-company operating system. Reducto still meets as a full team every week so engineering understands important deals and the business understands technical work. Some projects now require longer than a week, and the priorities have shifted toward monthly and quarterly horizons. The company had not adopted a fixed annual set of OKRs at the time of the interview, Abraham said, in part because the market changes quickly and priorities need to respond to new information.

The co-founders were described as having never had what Abraham would classify as an argument, though they had disagreed and talked through decisions. The important feature of their relationship, he said, is that they can assess disputes on their merits and resolve them without resentment. Their responsibilities are complementary: Chowdhuri leads much of the engineering work, while Abraham leads go-to-market. The ability to get through difficult periods without becoming resentful matters more than any formula for how co-founders should work together.

Reducto processed a little over a billion pages each month at the time of the interview, according to Abraham, and the volume was changing quickly. Continuous evaluation and a broad customer base are part of how the company encounters a wider range of documents and failure patterns. Abraham said the broader the customer base gets, the better the company’s intuition becomes about the range of problems it needs to handle.

He characterized competition as “meritocratic”: customers will look elsewhere if Reducto stops performing well. AI-native companies such as Harvey were named as customers who will seek out the strongest available solution. Reducto’s answer is to keep improving. Abraham argued that competitive advantage in this market depends less on a permanent systemic moat than on consistent execution.

The effects of document quality can extend beyond an extraction score. One example Abraham gave was a legal-technology customer that found users interacted with documents twice as much when those documents were parsed by Reducto. The customer attributed the difference to better answers and more precise citations, which made people more likely to ask questions. That kind of downstream impact matters even if it does not appear immediately in an evaluation.

At the time of the interview, Reducto was hiring across the business. Abraham said the company had about 40 roles open, including engineering, infrastructure, product, machine-learning, and platform positions. On the commercial side, the sales organization had grown from one account executive at the start of the year to 30 people. The company continued to feel capacity constrained, and hiring was not expected to slow in the near term.

Reducto’s product scope remains bounded by the document itself. Storage is less interesting than actions such as classifying, parsing, and editing a document. The standard is not to expand into every adjacent service; it is whether the company can do a task exceptionally well and whether doing it improves the experience for customers.