Kushal Byatnal spends his days on a problem most people would rather ignore: the billion-plus PDFs the world creates every day, and the fact that software still can't read them reliably. His company, Extend, exists to fix that.
Extend is a New York company that turns unstructured documents into production-ready, structured data. Byatnal is its co-founder and CEO, running the business alongside co-founder Eli Badgio. The pitch is simple to say and hard to build: take the hardest documents a business touches - tables, handwriting, signatures, multi-page packages - and pull clean, validated data out of them with more than 95% accuracy, fast enough to ship in days rather than months.
The customer list reads like a fintech roll call. Brex, Chime, Square, Mercury, Opendoor, Checkr, Flatiron Health, Vendr, and Tesorio all run documents through Extend. Some use it for back-office operations, fraud checks, and finance workflows. Others embed it directly in their own products, so an end user can upload a document and have it classified and validated in real time. Extend says its platform processes millions of pages a day.
That framing is the heart of Byatnal's argument. Storage got a cloud. Compute got a cloud. Documents - the data locked inside every scan, form, and email - never did. Extend is his attempt to build the missing layer.
Two startups, one instinct // track record
Byatnal is a two-time technical founder, and the through-line across his career is a dislike of tedious, manual work. He studied computer science at Duke University, graduating in 2017. Along the way he interned at Amazon Robotics, applying computer vision to make fulfillment centers more efficient, then worked as a software engineer at Google on natural language processing and query understanding. Vision and language, the two ingredients that now power Extend, were both on his resume years before he needed them.
He joined Brex as an early engineer, where he spent his time building internal tooling for support and operations teams. That experience left a mark. His earliest startup launches were explicitly about killing the "copy/pasting and context switching" he watched those teams suffer through every day. After Brex he co-founded Stir, a creator fintech company that grew to hundreds of millions in payment volume.
Extend started in 2023 in Y Combinator's Winter batch. The first versions were about internal tools and AI-powered workflows for unstructured data. Over time the company sharpened its focus onto the part customers found hardest: getting document data accurate enough to trust in production.
Why OCR was never enough // the product
There is no shortage of OCR and document-parsing tools on the market. Byatnal's team noticed that customers still struggled once accuracy requirements got high. Parsing, it turned out, was only one part of the problem. The real gap was between a raw parsed output and data clean enough to run a mission-critical pipeline on.
Extend's answer was to unify the whole stack - vision models, OCR, infrastructure, and tooling - into a single platform. Documents get parsed with specialized vision models, data gets captured through multi-step extraction with semantic chunking and bounding-box citations, and a human-in-the-loop review system catches the edge cases. Corrections feed back into fine-tuning pipelines that produce custom models, so accuracy improves over time instead of plateauing.
Illustrative comparison. Extend positions its platform around a >95% accuracy bar on customers' hardest documents.
In 2025 the company shipped Parse 2.0, which Byatnal called the most accurate document parsing API in the world, and open-sourced Extend UI - 14 MIT-licensed components for building document agents, including PDF, DOCX, and XLSX viewers plus bounding-box citations. The open-source move came from a familiar frustration: when the team started, they tried every existing file viewer, found them all lacking, and decided to build and give away their own.
The bakeoff test // what customers say
Byatnal tends to let results do the talking, and Extend's customers talk about accuracy in blunt terms. Vendr's CTO said the team ran a bakeoff and Extend had the best results of any solution on the market, eliminating "an entire class of engineering problems." A staff engineer at Checkr framed the value as offloading the maintenance cost of model tuning, scoring, and evaluations. Tesorio's CTO put it more plainly: he didn't know what Extend was doing under the hood, only that it was more accurate than anything he'd tried.
That reputation is what pulled in Extend's investors. The $17 million round, announced in June 2025, was led by Innovation Endeavors, with participation from Y Combinator, Homebrew, and Character. The angel list included Scott Belsky, the former chief strategy officer of Adobe, and Guillermo Rauch, the CEO of Vercel - both people who know a lot about developer tools and design.
What he's betting on // the long game
Byatnal's ambition is to make Extend the definitive cloud platform for document processing - the place teams reach for whenever they need to turn documents into reliable data, the same way they reach for a storage or compute provider today. It is an unglamorous bet in an industry that loves flashy demos. Documents are boring. That, in his telling, is exactly the point: the boring problem sits underneath a huge amount of real economic activity, and solving it well is worth more than chasing the hype cycle.
His career keeps circling the same idea. From Amazon Robotics to Google to Brex to Stir to Extend, the constant is a builder trying to remove the tedious layer between people and the work they actually want to do. Extend is the fullest expression of it so far - and, if the customer roster is any signal, the one that has stuck.