Extend: Document Processing APIs for Engineering Teams

Extend is a New York company, founded in 2023 in Y Combinator's Winter 2023 batch, that sells an API and dashboard for turning PDFs and scans into structured data. Its products parse documents, extract fields into a customer-defined schema, split bundled files, classify documents and fill PDF forms, and a workflow layer chains these steps. The buyer is an engineering team that wants to build document pipelines on top of it.

Company and ownership

Extend was founded by Kushal Byatnal and Eli Badgio. Byatnal was an early engineer at Brex and co-founded the creator payments company Stir, according to the Y Combinator profile. A $17M seed and Series A was announced in June 2025, led by Innovation Endeavors with Y Combinator, Homebrew and Character. Extend is private and lists no parent company. The Y Combinator profile puts the team at 25 people; the company does not publish a current headcount.

Where Extend fits

Extend's case studies describe three situations in which teams chose it:

• Several tools to replace. Brex says document processing was split across four vendors before it moved to Extend and compared vendors on an evaluation set for accuracy, latency, cost and customization.
• Latency limits. Mercury tested AWS Textract, foundation models and other startups for onboarding documents. It chose Extend for a fast mode with a p90 latency under 7 seconds and for bounding boxes that cite where each value came from.
• Extraction built in-house. Flatiron Health's machine learning team had tried to put oncology report extraction into production from 2018 to 2023, and says Extend replicated six months of that work in two weeks.

The named industries are fintech and lending (Brex, Mercury, Checkr), healthcare (Flatiron Health), real estate (HomeLight, Opendoor) and freight (Nuvocargo, listed in a case study). Each case study is published by Extend. The company also names Square, Amgen, FactSet and C.H. Robinson on its homepage, with no case study behind those logos.

Extend is a poor fit for teams that need a general document repository or an application for business users. It is built for developers: the dashboard and its Studio exist for configuration and review, and the integration runs through an API, one of four SDKs (Python, TypeScript, Java, Go), a CLI or an MCP server. A one-off OCR job without a schema or evaluation set uses little of what Extend sells.

Products

Around the products sit evaluation sets with accuracy metrics, versioned processors, pinned base model versions, and Composer, an agent that tunes extraction schemas. Parse 2.0 runs a layout model first and then specialized models per block type, and was trained on 1M+ real-world pages.

Click to load video from YouTube

Product video on Parse 2.0, from the Extend YouTube channel. The channel's newest upload, a panel with Brex, Baseten and Vercel from October 2026, is a discussion rather than a product demonstration.

Deployment, security and pricing

Extend runs as managed cloud in the US (us1, us2) or the EU (eu1, Frankfurt), as BYOC in the customer's own cloud account, or as hybrid, in which documents stay in the customer's cloud and inference runs at Extend. BYOC needs GPU quota and about two to three weeks of provisioning. There is no on-premises or air-gapped option. The documentation says foundation-model providers work under zero-retention terms with no training on customer data, and that zero data retention is available on request.

On compliance, Extend claims HIPAA (with business associate agreements) and GDPR. For SOC 2 the documentation says only that controls are "aligned with SOC 2 Type II requirements" and points to a trust center; the wording does not state that a Type II report has been issued, so ask for the report.

Pricing is credit-based. Pay As You Go is free for the first 10,000 credits, then $0.0125 per credit. Scale costs $500 a month for 50,000 credits; Enterprise, which includes BYOC and hybrid, is custom. Extraction costs 3 credits per page plus 2 for parsing, so a 10-page document costs about $0.63 on Pay As You Go.

Evidence

All outcomes below come from Extend's own case studies.

Extend's benchmark claims are also self-published. RealDoc-Bench, which Extend wrote, reports 95.7% question-answering accuracy for Performance Parse, but the launch post names no competitor scores, and a later Auto-Router post gives 90.9% for the same mode. The Operator-1 announcement reports 99.3% on Extend's own LongArray-Extract and 99.88% on LongExtractionBench, a third-party benchmark co-designed by Reducto, where Reducto Deep Extract follows at 99.3%; all scores are as reported by Extend. In a competitor-run test, DocuPipe scored Extend at 80.28% against its own 97.02% on 72 public documents, and 95.15% against 95.31% on RealDoc-Bench; DocuPipe built that test, so it is no more neutral than Extend's.

Market position

Extend competes with Reducto, LlamaParse, Unstructured and DocuPipe among developer-first parsing and extraction APIs, and with cloud services such as Google Document AI and Textract. Its line has grown from parsing and extraction to parsing tiers (Light Parse in August 2026, Auto-Router in September) and, since October 2026, the Operator-1 extraction agent. Cloud providers keep adding agentic features to their own document APIs, so the lead Extend claims on hard documents depends on benchmarks that buyers should rerun on their own files. The wider IDP vendor directory lists comparable platforms.

Resources

• Website and documentation
• Pricing and case studies
• Deployment options and compliance
• Y Combinator profile and Series A announcement
• YouTube channel