Adit Abraham

Adit Abraham is the co-founder and CEO of Reducto, which builds infrastructure for AI systems working with complex documents. His work addresses a practical obstacle to automation: a model can reason fluently while misreading the table, handwritten note, or financial statement on which its decision depends.

From product management to document intelligence

Abraham studied computer science at MIT and conducted machine-learning research at the MIT Media Lab. His early entrepreneurial projects included Sidewalk, a Shopify plugin that recommended products from other stores during checkout. He met his future co-founder, Raunak Chowdhuri, in an MIT graduate machine-learning course. Their eventual partnership combined Abraham’s focus on product and customer development with Chowdhuri’s technical focus, as described in Reducto’s path to product-market fit.

Before Reducto, Abraham worked as a product manager at Google on YouTube advertising and Search. He enjoyed advertising’s clear business targets and measurable outcomes; his subsequent Search work involved more coordination, with impact that felt less immediate. His reflection on leaving Google argues for choosing career opportunities around learning, consequential responsibility, and the work itself, whether inside a large company or a startup.

Abraham and Chowdhuri founded Reducto in 2023 and joined Y Combinator’s Winter 2024 batch. Their original product, Remembrall, gave language models long-term memory. Customer conversations exposed a more urgent need: getting usable information out of uploaded files. The founders had encountered that difficulty in their own application, and a document-processing experiment drew stronger demand, prompting their pivot to parsing.

An early playground let prospective customers upload documents that had defeated other tools. It established a practical sales approach: demonstrate the improvement on the customer’s own material. Abraham initially focused on startups in finance, insurance, and legal work, where accuracy mattered and teams could integrate quickly. Manual onboarding kept him close to failures and customer requirements. His account of founder-led customer development explains why he retained direct involvement in sales and customer success as sources of product knowledge.

Enterprise adoption required a different kind of persistence. Reducto’s first Fortune 10 sales process lasted 154 days and involved technical evaluations, acquisition discussions, procurement negotiations, and an air-gapped deployment. The founders’ account of that deal emphasizes an internal customer champion and continued delivery after the contract was signed. Demonstrating accuracy was only part of earning and keeping the customer.

Faithful reading, useful representations

Abraham connects document processing to the decisions that follow it. As AI applications move from answering questions to taking multiple actions, he argues that ingestion errors can propagate through later decisions. His approach to document intelligence combines several distinct priorities:

• Faithful document recovery: Correcting a reading error must preserve what the document actually says. A system should distinguish a zero from the letter O while retaining a printed total even if its arithmetic is wrong. Reducto’s agentic OCR adds targeted verification and correction after an initial parse, reducing the risk that a model regenerates the text and silently substitutes more plausible content.

• Representations suited to their consumers: Markdown can represent simple tables efficiently; HTML can preserve relationships encoded by merged cells in complex tables. Retrieval has a different need: an embedding model may struggle to match a question about changing revenue to tags and numbers. Abraham advocates retaining the structured table for reasoning and creating a separate natural-language representation for retrieval.

• Complementary vision tools: Abraham favors efficient computer vision for layout detection and vision-language models for content requiring semantic interpretation, including handwriting. This division accounts for processing cost, predictability, and the difficulty of particular document regions. For chart extraction, he describes Reducto’s team using agents with code interpreters, plotting, and repeated checks to reconstruct data tables. The team’s chart-extraction research also explains the limits of visual verification and its use of specialized models for axes, legends, and data series.

• Evaluation throughout the workflow: Parsing, retrieval, formatting, and final agent performance can fail independently. Accurate transcription offers little help when retrieval supplies the wrong passage. Abraham advocates evaluating each stage and monitoring production data, which can differ from benchmark examples. Reducto’s RD-TableBench contributes an open evaluation resource with 1,000 manually annotated examples of difficult tables.

He also treats classification and splitting as important parts of document orchestration: identifying document types and routing relevant sections lets models concentrate on the material needed for a task. His work extends that idea to file-system and CLI interfaces through which agents can choose document tools as needed, rather than sending every file through one predetermined sequence. These interfaces distinguish readable content from metadata, including bounding boxes useful for citations. Document orchestration and agent interfaces.

Expanding the document workflow

Under Abraham and Chowdhuri, Reducto expanded from parsing into extraction, splitting, classification, editing, and tools for document workflows. The founders announced a $75 million Series B in October 2025, bringing total funding to $108 million and supporting that broader direction.

In September 2026, Reducto introduced r-1 in preview, a parsing model designed to handle layout, tables, formatting, and citations in a single pass rather than through multiple model calls. Across these approaches, Abraham’s focus is giving AI systems a faithful, usable account of the documents people already work with—and evaluating whether that account improves the actions those systems take.