Reducto

MIT graduates Adit Abraham and Raunak Chowdhuri founded Reducto in 2023 after a weekend project exposed how poorly conventional OCR and general-purpose language models handled real enterprise documents. Abraham had worked in Google Ads/Search product management and MIT Media Lab machine-learning research; Chowdhuri had published computer-vision work from an early age and worked in MIT research and applied ML. They entered YC W24 and built an API-first ingestion layer that combines traditional OCR/computer vision, proprietary models and frontier vision-language models rather than relying on a single model. Reducto's Parse API preserves text, layout, tables, equations, handwriting, images and bounding boxes across 30+ file types and 100+ languages; Split and Classify segment mixed packets; Extract maps content to requested schemas with citations; Edit modifies source documents; Studio supports testing/evaluation; pipelines, webhooks, CLI and MCP connect the functions. Agentic OCR and Deep Extract add iterative review/correction loops intended to catch last-mile mistakes on long, dense documents. The business targets AI-native application companies and regulated enterprises that would otherwise maintain brittle OCR, chunking and extraction pipelines. Deployments range from multi-tenant cloud to customer VPC, on-premise and fully air-gapped environments, with SOC 2 Type II, HIPAA/BAA, data residency and zero-data-retention controls on higher tiers. Reducto raised an $8.4m seed led by First Round in October 2024, a $24.5m Series A led by Benchmark in April 2025 and a $75m Series B led by Andreessen Horowitz in October 2025, totaling $107.9m mathematically and $108m as announced. Other backers include Y Combinator, BoxGroup, SV Angel, Liquid 2, WndrCo and operator angels. The Series B valued it at roughly $600m according to Forbes. Processing crossed 250m pages by April 2025 and 1b pages by October 2025; 2026 company materials claim more than 4b cumulative pages and capacity of hundreds of millions per day. Named customers include Harvey, Scale AI, Vanta, Mercor, Rogo, Airtable, Anterior, August, LEA and Elysian, plus unnamed Fortune 10 and top-five hedge-fund deployments. Public Standard pricing includes the first 15,000 credits free and then $0.015/credit; a standard parse is one credit/page and batch parsing is 20% cheaper. Growth and Enterprise are negotiated. Competitive and diligence risks include rapid commoditization by frontier multimodal models, document heterogeneity, benchmark sponsorship bias, high GPU/inference costs, sensitive-data exposure, customer concentration in AI startups, and strong rivals including Google/AWS/Azure document AI, LlamaParse, Unstructured, Extend, LandingAI, Instabase, Rossum and Hyperscience.

Founding story

A weekend document-processing hack by two MIT friends exposed the accuracy gap between classic OCR and what AI applications needed. They named Reducto after the Harry Potter shattering spell, joined YC W24 and focused on converting messy human documents into faithful machine-usable structures.

Business model

Usage-based document-intelligence API and developer platform, with enterprise subscriptions and private/on-premise deployment.

Free initial credits followed by per-credit usage; negotiated Growth volume commitments; Enterprise contracts for dedicated throughput, support, security, SLA and private deployment.

Traction

Processed 250m+ cumulative pages by April 2025, more than 1b by October 2025 and claims 4b+ by July 2026; monthly processing grew 6x in the six months before Series B. Current materials claim hundreds of millions of pages/day capacity and 99.9%+ uptime. YC lists 30 employees. Named users include Harvey, Scale AI, Vanta, Mercor, Rogo and Airtable.

Latest developments

Deep Extract launched April 2026 and ranked first in the Reducto-commissioned, micro1-published LongExtractBench in June 2026. Current 2026 platform adds MCP/agent interfaces, Edit, Studio evaluations, regional/private deployment and claims more than 4b pages processed.