San Francisco startup Unsiloed AI has raised $7.5 million in seed funding led by Felicis to help companies feed complex documents to AI agents without errors. Its software breaks down PDFs with tables, charts and complicated layouts, extracts the content as structured data, and shows exactly where each value came from in the original file.
Quick Take
- The round: $7.5 million seed, led by Felicis, with Y Combinator, Orange Collective and Uncorrelated Ventures, plus angels Aarthi Ramamurthy and Beth Turner.
- The product: Document parsing that turns tables, charts and layouts into structured data for AI agents, with links back to the source location.
- The claim: 99.78% accuracy on about 84,000 data points from 542 financial documents of one customer, according to the company’s own technical brief. Not independently verified.
- The price: From one cent per page, the company says.
- The team: Seven people, founded by Aman Mishra and Adnan Abbas, Y Combinator Fall 2025 batch.
In plain English
AI agents are good at reading plain text but often stumble over a table in a bank statement or a chart in a contract. A wrong number pulled from a document can cause expensive mistakes. Unsiloed reads those documents first, turns the hard parts into clean data, and marks where every value came from so a person can check it.
What happened
Unsiloed announced the seed round on November 18. The company says it already works with customers in finance, legal services and healthcare, including a bank that processes documents with it and a law firm that runs it on its own servers. It names voice AI startup Bolna AI as a customer.
Technically, Unsiloed says it uses “dual-stream” vision models that process text tokens and image tokens separately, plus what it calls heat map based chunking, which puts more computing power on the parts of a page that matter most, such as figures and numbers.
Background
Document parsing has become a busy market as companies connect agents to internal files. Unsiloed competes with specialised startups such as Reducto and Unstructured as well as offerings from Microsoft and Amazon. On olmOCR-Bench, a public benchmark from the Allen Institute for AI covering 1,403 PDF pages, Unsiloed reports a score of 88.0 in its own tests, against 66.0 for Reducto’s agentic system and 39.9 for Unstructured. The money will go to hiring researchers and engineers, sales and proprietary datasets.
Why it matters
Most enterprise knowledge sits in PDFs, scans and spreadsheets, not in clean databases. As AI agents take on real work in banks, hospitals and law firms, the quality of the data they read becomes a hard limit on how far they can be trusted. Traceable extraction is one way to keep humans able to check the output.
What’s still open
- Independent tests of the accuracy and benchmark claims.
- Names of the bank and law firm customers.
- How the approach holds up against general-purpose models that keep getting better at reading documents.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent i10x analysis is based on Unsiloed AI’s funding announcement and statements by the company. Benchmark figures are the company’s own results.
Sources (plain text, no links):
- Unsiloed AI: announcement “Unsiloed raises $7.5M seed round led by Felicis Ventures” (November 18, 2026)
- Unsiloed AI technical brief on extraction accuracy (company provided)
- Allen Institute for AI: olmOCR-Bench (public benchmark)
i10x Perspective
Small rounds like this one show where the agent stack is still thin. Models get the headlines, but agents fail quietly on a misread table. Tools that make documents reliable could be among the most useful plumbing of the agent era. More startup funding on our Startups page and in the AI Funding Index.



