Unsiloed AI has raised $7.5 million in seed funding to develop software that converts documents into structured data for applications and AI agents. The startup’s system processes files such as PDFs, scanned forms, presentations, Word documents and images, returning results in formats including Markdown and JSON through API calls.
The company is focused on a limitation of conventional optical character recognition: extracting words alone may strip away the layout and relationships that give a document meaning. A figure’s position in a table, the heading associated with a contract clause, or the axes on a chart can be as important as the characters themselves. Unsiloed says its approach analyzes text and visual structure together.
Parsing content and layout together
Unsiloed’s software uses what the company describes as a dual-stream vision architecture. It is designed to process document content alongside its visual organization, rather than treating a page as an undifferentiated sequence of text. Developers can use the system to extract fields according to custom schemas, classify and split documents, and update text or images while preserving formatting.
The target material includes financial statements, insurance claims, contracts, clinical records and operational forms. These documents often contain information that businesses need to make available to software systems, but their varied formats can make it difficult to use that information consistently. Structured outputs can give downstream applications a more defined representation to work with, although the quality of the result still depends on the parsing process and the document being processed.
Unsiloed also provides confidence scores and citations to the source material, which the company says can help teams validate extracted information and send uncertain results for review. The controls are relevant because the system is not described as perfectly accurate. In settings where a mistaken field could affect a business process, review mechanisms can help users distinguish results that may need additional scrutiny.
Benchmark result and deployment choices
Co-founder and CTO Adnan Abbas said Unsiloed’s parser ranked first among 19 systems in the company’s published comparison on olmOCR-Bench, with a score of 88.0. The result was calculated using Allen Institute for AI (Ai2)’s unmodified deterministic scorer, according to the company. The benchmark tests document-parsing tasks that include complex tables, mathematical expressions, multi-column reading order and scanned pages.
Ai2 created olmOCR-Bench to assess how well vision-language models and OCR systems turn PDF page images into structured text. The suite contains roughly 1,400 PDF pages and more than 7,000 binary unit tests covering layout, reading order, mathematical rendering and table structure. The score is a result within that benchmark comparison; it does not by itself establish performance across every document type or customer environment.
Unsiloed says customers can deploy its product in private clouds, on premises or in air-gapped environments. Those options are intended to let organizations process sensitive documents within their own infrastructure. The company has not provided additional deployment details in the information about the funding announcement.
Funding to support product and deployment work
The seed round was led by Felicis. Participants included Y Combinator, Orange Collective, Uncorrelated Ventures, Aarthi Ramamurthy and Beth Turner, among others. Unsiloed said it will use the funding to develop its document models, expand its engineering team and grow cloud and on-premises deployments for enterprise customers.
Co-founders Aman Mishra and Adnan Abbas started Unsiloed in the 2024–2025 period, and the company joined Y Combinator’s Fall 2025 batch. Mishra, the company’s co-founder and CEO, described the product as an interface for AI agents to read, update and act on unstructured data. That ambition places document parsing between source files and the systems expected to use their contents: the practical challenge is not only recognizing text, but retaining enough structure to make the extracted information useful and reviewable.