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

1 hour ago 3

As businesses progressively crook to AI agents to automate work, the immense amounts of accusation locked successful analyzable documents stay a challenge. Unsiloed AI has raised $7.5 cardinal successful effect backing led by Felicis to tackle that problem, aiming to assistance enterprises marque their documents much accessible to AI agents done specialized technology.

Silicon Valley task superior steadfast Felicis led the round, with information from Y Combinator, Orange Collective, and Uncorrelated Ventures. Angel investors Aarthi Ramamurthy and Beth Turner besides participated.

Business documents often see tables, charts, and analyzable layouts that AI systems tin conflict to construe accurately. But getting the accusation incorrect tin pb to costly mistakes for businesses, and the interaction could beryllium terrible successful the lawsuit of a bank, hospital, oregon a instrumentality firm, arsenic they trust connected these documents for their regular work.

Unsiloed aims to lick this occupation with bundle that breaks down analyzable documents, extracts accusation from tables, charts and different elements, and turns it into structured information that AI agents tin use. The exertion besides shows wherever the extracted accusation came from successful the archetypal document, truthful businesses tin cheque the results alternatively than instrumentality the AI’s output astatine look value.

“The occupation with LLMs is they’re inherently probabilistic,” said Unsiloed co-founder and CEO Aman Mishra (pictured above, right) successful an interview. “There is going to beryllium immoderate deviation induced. There is going to beryllium immoderate hallucination involved, and that has a monolithic antagonistic interaction downstream.”

Mishra co-founded Unsiloed with Adnan Abbas (pictured above, left), his erstwhile classmate astatine the Indian Institute of Technology (IIT) Kharagpur. The 2 joined Y Combinator’s Fall 2025 batch aft spending immoderate clip processing the technology.

Unsiloed is already moving with customers successful finance, ineligible services, and healthcare, Mishra told TechCrunch. One slope is utilizing its exertion to process documents, portion a instrumentality steadfast has deployed it connected its ain servers to grip contracts and aged lawsuit files, helium said, without naming either customer. Mishra, however, named dependable AI startup Bolna AI arsenic a customer, saying it uses Unsiloed to ingest cognition bases for gathering dependable agents.

The San Francisco-headquartered startup has entered a marketplace that is already crowded and includes specialized document-processing startups specified arsenic Reducto and Unstructured, arsenic good arsenic offerings from Microsoft and Amazon. Businesses besides sometimes usage general-purpose AI models to extract accusation from documents, giving Unsiloed different acceptable of competitors.

Mishra told TechCrunch that Unsiloed’s vantage lies successful however it processes analyzable documents. The startup uses what it calls “dual-stream” imaginativeness models, which grip substance and representation information separately, on with a method called “heat map-based chunking” that helps place the astir important parts of a document.

“In 1 of the streams you person the substance tokens, successful the different watercourse you person the nonstop representation tokens,” Mishra said. “This is thing that we person spent a batch of clip architecting.”

The exertion past directs much computing powerfulness toward areas specified arsenic figures, numbers, and charts. Unsiloed besides uses a specialized decoder that takes industry-specific terminology into account, helping it separate betwixt presumption that tin mean antithetic things successful antithetic concern contexts.

Supporting those claims, Mishra pointed to a trial involving 542 fiscal documents supplied by an endeavor customer. Unsiloed achieved 99.78% accuracy crossed astir 84,000 extracted information points, per a method little shared by the startup. The trial compared Unsiloed with large AI models, Microsoft’s Azure Document Intelligence, and the customer’s in-house extraction tool. The results have, however, not been independently verified.

Unsiloed besides claims to person outperformed competing AI document-processing tools connected olmOCR-Bench, a nationalist benchmark developed by Allen AI. In tests conducted by the startup, Unsiloed scored 88.0, compared with 66.0 for Reducto’s agentic strategy and 39.9 for Unstructured. The benchmark covered 1,403 PDF pages and tested however good the systems handled tables, mathematical expressions, scanned documents, and analyzable layouts.

Unsiloed’s pricing, Mishra said, starts astatine 1 cent per page. He asserted that the startup’s attack besides helps support its costs down by directing computing powerfulness toward the parts of a papers that request the astir attention, alternatively of processing each conception successful the aforesaid way.

Currently, Unsiloed has a headcount of 7 people. The startup, Mishra said, plans to usage the caller backing to prosecute much researchers and engineers, grow its income efforts and make proprietary datasets to amended its technology.

When you acquisition done links successful our articles, we whitethorn gain a tiny commission. This doesn’t impact our editorial independence.

Read Entire Article