Extend

Building @extendHQ

Extend is a document processing platform for AI teams that need to turn messy PDFs into reliable structured context for agents and pipelines. With this launch, the New York based company is releasing Parse 2.0, its document parsing API built for agents and pipelines where accuracy, cost, and latency are critical, alongside RealDoc-Bench, a companion benchmark that measures parsing performance on real-world documents from logistics, healthcare, financial services, and real estate. The launch matters because document parsing is still a bottleneck for production agents working with the hospital intake forms, mortgage packets, tax filings, and bills of lading that businesses actually run on. Rather than relying on a single large model to read a page end to end, Parse 2.0 uses a layout-first approach that segments each page into regions and routes those regions to a suite of fine-tuned OCR and vision language models downstream. On RealDoc-Bench, Extend reports an adjusted F1 of 0.847 on layout accuracy across 1,500 samples, ahead of Reducto at 0.759, AWS Textract at 0.709, Azure DI at 0.687, and PaddleOCR-VL at 0.684. On document Q&A accuracy across 1,359 prompts and 581 documents, Parse 2.0 scored 95.7%, with LlamaParse (Agentic) at 92.1%, Reducto (Agentic) at 91.1%, and Extend Parse 1.0 at 90.4%. Founded in 2023 by Eli Badgio and Kushal Byatnal, Extend is based in New York and counts Brex, Mercury, and Opendoor among the teams running documents through its API. For founders and operators building agents that touch regulated or systems-of-record documents, Parse 2.0 and the public RealDoc-Bench results give a concrete way to compare parsing options on workloads that look like their own, instead of clean academic PDFs.

Tags

Comments (9)