Felicis leads $7.5M bet on Unsiloed AI to make enterprise documents usable by AI agents

Startups & Funding

Unsiloed AI raises $7.5M to turn documents into agent-ready data

Unsiloed AI raised $7.5 million led by Felicis. The seven-person startup says its document extraction reached 99.78% accuracy, but its benchmark results remain unverified.

By
Sophie Lindqvist
Filed
Channel
Startups & Funding
Read
4 min read

By Sophie Lindqvist

Unsiloed AI has raised a $7.5 million seed round led by Felicis to build software that converts complex enterprise documents into structured data for AI agents.

The San Francisco startup had seven employees when co-founder and CEO Aman Mishra described the round to TechCrunch. Silicon Valley venture capital firm Felicis led the investment, with Y Combinator, Orange Collective and Uncorrelated Ventures participating. Angel investors Aarthi Ramamurthy and Beth Turner also joined.

Why do complex documents still block AI agents?

Enterprise records often combine prose, tables, charts and difficult layouts, according to Unsiloed. The startup says AI systems struggle to interpret those elements accurately. It argues that extraction errors can become costly in banks, hospitals and law firms, where employees depend on those documents for daily work.

Unsiloed’s software breaks documents into parts, extracts information from tables, charts and other elements, and converts it into structured data for agents. It also identifies where each extracted item came from in the original document, so customers can inspect results rather than accept model output without review.

Mishra said: “The problem with LLMs is they’re inherently probabilistic. There is going to be some deviation induced. There is going to be some hallucination involved, and that has a massive negative impact downstream.”

Who is using Unsiloed?

Unsiloed already works with customers in finance, legal services and healthcare, Mishra told TechCrunch. One unnamed bank uses its technology to process documents, Mishra said, while an unnamed law firm has deployed it on its own servers for contracts and old case files.

Bolna AI, a voice AI startup, uses Unsiloed to ingest knowledge bases for building voice agents, according to Mishra. The bank and law firm remained unnamed. Unsiloed did not disclose contract values, deployment volumes or revenue.

Mishra co-founded the company with Adnan Abbas, his former classmate at the Indian Institute of Technology Kharagpur. The pair joined Y Combinator’s Fall 2025 batch after spending time developing the technology.

What does the dual-stream system do?

Unsiloed uses what it calls “dual-stream” vision models, which process text and image data separately, Mishra said. Its “heat map-based chunking” identifies important parts of a document, while additional computing power goes to figures, numbers and charts.

“In one of the streams you have the text tokens, in the other stream you have the exact image tokens,” Mishra said. “This is something that we have spent a lot of time architecting.”

A specialized decoder also incorporates industry terminology, helping distinguish words that carry different meanings in separate business contexts. Unsiloed says selective processing lowers its costs, but it gave no figure for savings or latency. The company also did not disclose model size, training compute or decoder architecture.

How strong are the accuracy claims?

Unsiloed reported 99.78% accuracy across roughly 84,000 extracted data points from 542 financial documents supplied by an enterprise customer, in a technical brief shared by the startup.

The startup’s test compared Unsiloed with major AI models, Microsoft’s Azure Document Intelligence and the customer’s in-house extraction tool. No independent party has verified the results.

On olmOCR-Bench, a public benchmark developed by Allen AI, Unsiloed scored 88.0 in tests run by the startup. In those tests, Reducto’s agentic system scored 66.0, and Unstructured scored 39.9.

The benchmark covered 1,403 PDF pages and tested tables, mathematical expressions, scanned documents and complex layouts. Unsiloed did not disclose the full test protocol or provide independent confirmation of the three scores. It reports 88.0 as a benchmark score, not as a percentage accuracy rate.

Can one-cent pricing win enterprise deployments?

Unsiloed competes in a crowded market with specialized document-processing startups Reducto and Unstructured, plus offerings from Microsoft and Amazon. Businesses also turn to general-purpose AI models for document extraction. Its starting price is one cent per page, Mishra said.

Mishra said Unsiloed’s advantage lies in its handling of complex documents. The system combines separate text and image streams, traces extracted data and directs compute selectively. Unsiloed did not provide customer contract terms, measured savings or competitor pricing.

Mishra said Unsiloed plans to use the $7.5 million seed round to hire researchers and engineers, expand sales efforts and develop proprietary datasets. With seven staff now and new capital aimed at those tasks, the next test is whether benchmark leads become paid volume while the one-cent-per-page entry price holds.

Source: techcrunch.com

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,125 articles

Related articles

« Previous article