Adit Abraham on Why Document Parsing Still Breaks AI

Allen Park welcomes Reducto co-founder and CEO Adit Abraham to Latent Space Cooking. While preparing a steak sandwich with chimichurri, burrata and peaches, they discuss why document processing is difficult and how the startup approaches it.

Abraham says the founders entered Y Combinator with Remembrall, an agent-memory idea, but already expected to pivot. The Reducto name came from a group discussion about potential names and available domains.

PDF processing has a long history, but its original goal was faithful printing rather than making information easy for AI agents to read. Tables, layout and visual relationships make reliable extraction a difficult vision problem.

Abraham compares incremental document recognition to driver assistance and the ambition of reliable end-to-end reading to autonomous driving. Reducto wants models to interact with documents beyond parsing, including classification and editing.

The long tail complicates routing: an ordinary-looking document can contain one difficult table whose location is unknown in advance. The company initially used smaller computer-vision models for routine work and larger models for difficult cases.

Abraham introduces R1 as Reducto’s in-house model for end-to-end processing, including layout and corrections. He frames model choice around accuracy, cost and latency rather than a commitment to building everything internally.

He says task-specific post-training has often improved the company’s cost, accuracy and latency trade-off, while general-purpose foundation models still make sense for some broader tasks. The appropriate split depends on the workload.

Discussing Deep Extract, Abraham says frontier models can silently miss rows or context even when their extracted fields look precise. Missing information is especially hard to notice in production because the user may not know what should have been extracted.

The host asks about benchmark inflation. Abraham says vendor-created benchmarks often become marketing material; Reducto worked with an outside partner so it would not choose documents or ground truth. He still urges customers to test on their own documents.

At filming, Abraham says Reducto processed more than a billion pages per month. He presents R1 as a way to improve accuracy while simplifying and reducing per-page pricing; these figures and comparisons are company claims.

Park asks why even 99% accuracy can be inadequate in healthcare or other high-stakes workflows. Abraham says first-pass accuracy matters, but no system is perfect and some mistakes have much greater consequences than others.

He emphasizes citations to source locations, reliable confidence estimates and escalation to a human reviewer. The team measures not only extraction accuracy but whether imperfect results can be identified and reviewed.

Reducto’s ML team conducts continuous evaluations on production failure patterns. A tilted faxed page with handwriting outside a grid illustrates how real cases can combine difficulties absent from a simple handwriting benchmark.

Abraham argues that documents in different industries share primitives: paragraph boundaries, reading order and table relationships. Improving resilience to messy healthcare documents may therefore help financial documents too, though domain-specific details remain.

He gives legal strikethroughs and underlines as examples of details that matter differently across domains. His goal is broadly applicable parsing, with configurable interpretation when a customer’s business definitions vary.

Asked about internal AI adoption, Abraham says the team did not mandate a particular coding agent. Employees saw colleagues use tools effectively and began applying them themselves; he mentions sales staff creating prospect-specific material.

The company allows language models during interviews and uses an eight-hour onsite. Abraham says the exercises are demanding enough that effective tool use can be an advantage, and recounts a candidate completing the software portion unusually quickly.

On hiring, Abraham says he looks for sustained attention to craft: a researcher who inspects individual PDF pages or an engineer who follows difficult customer-support problems through to resolution.

He asks candidates about something they have worked hard to master, including interests outside their job. His view is that curiosity and rigor can transfer when a person is matched with a problem they care about.

The conversation pauses for the peaches: the guests try a torch, then move the slices to a pan. They are building the sandwich alongside the interview rather than presenting a measured recipe.

As the company grew toward a Series B, Abraham says his work shifted from handling individual deals and support tickets to hiring, customers and communicating culture. Learning to delegate was uncomfortable at first.

The team still meets weekly, he says, but important projects can span more than a week. He now considers monthly and quarterly priorities while trying to preserve a startup’s ability to adjust as the market changes.

On competition, Abraham points to continuous evaluation and the variety of customers as sources of learning. He argues that demanding customers will switch if a supplier stops delivering the best results.

Reducto wants to focus on interactions with documents—reading, classifying and editing—rather than expanding into file storage for its own sake. Abraham says expansion should improve the customer experience.

He describes a legal-technology customer that reported more document engagement after improved parsing and citations. He treats the observation as a downstream customer example, not a universal benchmark.

Abraham and co-founder Raunak Chowdhuri met at MIT. He values their ability to discuss disagreements on their merits and work through startup pressures without building resentment; he describes their skills as complementary.

An Anthropic hackathon helped start their founding journey. Limited model context windows led them toward long-term memory before they pivoted to document processing.

Abraham says many products still involve search followed by reasoning, even if they do not look like a traditional retrieval-augmented chatbot. An agent navigating a file system also retrieves and combines relevant context.

The company is hiring across engineering, machine learning, infrastructure and sales. Abraham says growth has shifted the bottleneck toward having enough people to meet demand.

The conversation turns from reading to editing. Abraham wants models to handle tasks such as contract redlining and form filling, where an accurate understanding of structure must be followed by an equally careful modification.

Translation shows the challenge: changing an English document into Arabic may change text direction, layout and pagination. Good translated words alone do not ensure a usable edited document.

Abraham says the ML team works across problems rather than splitting completely by product. He values matching an engineer’s enthusiasm with a specific difficult problem, such as document editing.

He argues that office work has not yet had its own Claude Code moment. One obstacle is fragmented enterprise data and undocumented local procedures that a model needs in order to act reliably.

Specialist customers can understand the full workflow in a particular industry, while a document platform supplies a common capability. Abraham points to legal and tax applications as examples of this division of labor.

He says Reducto is improving structured extraction, citations and schema adherence while lowering cost. The price and performance comparison he gives reflects the company’s reported product changes at the time of recording.

The speakers assemble the sandwich with steak, burrata, peaches and chimichurri. They taste it and return to the playful question of whether Abraham’s earlier steak had been unfairly criticized online.

After a bite, they count the pieces and praise the sandwich. The episode closes with the host asking for a document-processing metaphor for the finished meal.