
Felicis Backs Unsiloed AI With $7.5M to Unlock Enterprise Documents for AI Agents
Author : Ravit Jain
The growing adoption of AI agents across enterprise workflows has exposed a persistent bottleneck: the enormous volume of business-critical information buried inside complex documents that AI systems struggle to interpret reliably. A startup called Unsiloed AI has secured $7.5 million in seed funding to address that gap, building technology designed to make dense, structured documents accessible and usable for AI-driven automation.
The seed round was led by Felicis, the Silicon Valley venture capital firm, with additional participation from Y Combinator, Orange Collective, and Uncorrelated Ventures. Angel investors Aarthi Ramamurthy and Beth Turner also joined the round.
The Problem With Documents and AI
Enterprise documents are rarely simple. Financial statements, legal contracts, medical records, and similar materials frequently contain tables, embedded charts, intricate formatting, and domain-specific terminology. These characteristics make it difficult for AI systems to extract information with the level of accuracy that businesses require. For organizations in sectors like banking, healthcare, and law, errors in document processing can trigger serious downstream consequences, whether in compliance, decision-making, or client outcomes.
Unsiloed's software is built to decompose these complex documents, pull out information from tables, charts, and other structural elements, and convert it into clean, structured data that AI agents can act on. Crucially, the system also traces extracted information back to its precise location in the source document, giving businesses a way to audit outputs rather than relying solely on what the AI produces.
"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."
Founding Team and Early Development
Unsiloed was founded by Aman Mishra and Adnan Abbas, former classmates at the Indian Institute of Technology (IIT) Kharagpur. The two spent time developing the core technology before joining Y Combinator's Fall 2025 batch, which provided additional resources and visibility as they prepared to bring the product to market.
The San Francisco-based startup is already working with paying customers across finance, legal services, and healthcare. Mishra confirmed that a bank is currently using the platform to process documents, and a law firm has deployed the system on its own servers to manage contracts and historical case files, though he declined to name either organization. He did, however, identify voice AI company Bolna AI as a named customer, noting that Bolna uses Unsiloed's technology to ingest knowledge bases for building voice agents.
A Crowded but Growing Market
Unsiloed enters a competitive landscape. The document processing and extraction space already includes dedicated startups such as Reducto and Unstructured, as well as established enterprise tools from Microsoft and Amazon. Some organizations also attempt to handle document extraction using general-purpose large language models, adding yet another layer of competitive pressure for the startup.
Mishra maintains that Unsiloed's technical approach sets it apart. The company uses what it describes as "dual-stream" vision models, which process text data and image data through separate channels simultaneously. This is combined with a proprietary technique called "heat map-based chunking," which identifies and prioritizes the most information-dense regions within a document.
"In one of the streams you have the text tokens, in the other stream you have the exact image tokens. This is something that we have spent a lot of time architecting."
Technical Architecture and Accuracy Claims
The dual-stream approach allows Unsiloed's system to allocate more computational resources toward the elements of a document that require the most precision - figures, numerical data, charts, and complex tables - rather than treating every section of a document with equal processing weight. The platform also incorporates a specialized decoder designed to handle industry-specific vocabulary, enabling it to distinguish between terms that carry different meanings depending on the business context in which they appear.
To support its accuracy claims, Mishra pointed to an internal evaluation involving 542 financial documents provided by an enterprise customer. Across approximately 84,000 extracted data points, Unsiloed reported an accuracy rate of 99.78%, according to a technical brief shared by the company. The evaluation compared Unsiloed's output against results from major AI models, Microsoft's Azure Document Intelligence service, and the customer's own in-house extraction tool. It is worth noting that these results have not been independently verified by a third party.
The company has also published benchmark comparisons on olmOCR-Bench, a public evaluation framework developed by the Allen Institute for AI. In tests conducted by Unsiloed itself, the platform scored 88.0 on the benchmark, compared to 66.0 for Reducto's agentic system and 39.9 for Unstructured. The benchmark covers 1,403 PDF pages and evaluates how well systems handle tables, mathematical expressions, scanned documents, and complex page layouts.
Pricing and Cost Efficiency
Unsiloed's pricing begins at one cent per page. Mishra argued that the startup's selective resource allocation model - directing compute toward the most demanding parts of a document rather than applying uniform processing throughout - also helps keep operating costs lower than competing approaches. This, he said, allows the company to offer competitive pricing while maintaining high accuracy on complex materials.
Plans for the Funding
Unsiloed currently operates with a team of seven people. According to Mishra, the $7.5 million in new capital will be directed toward hiring additional researchers and engineers, expanding the company's sales capacity, and building out proprietary datasets that will be used to further improve the accuracy and capabilities of its document processing technology. The startup's focus remains on serving industries where document accuracy is not just a technical requirement but a business-critical necessity.
Source Credits : TechCrunch


