Unsiloed AI, a San Francisco-based startup focused on enabling AI agents to accurately interpret complex business documents, has secured $7.5 million in seed funding. The round was led by Silicon Valley venture capital firm Felicis, with additional participation from Y Combinator, Orange Collective, Uncorrelated Ventures, and angel investors Aarthi Ramamurthy and Beth Turner.
The company aims to address a critical challenge in the enterprise AI sector: the reliable extraction of information from documents containing intricate tables, charts, and layouts. Co-founder and CEO Aman Mishra noted that large language models are inherently probabilistic, which can lead to hallucinations and significant downstream errors in critical sectors like banking, healthcare, and legal services. Unsiloed’s solution involves breaking down these complex files into structured data while tracing the source of extracted information to allow for human verification.
Mishra co-founded the startup with Adnan Abbas, his former classmate at the Indian Institute of Technology (IIT) Kharagpur. The duo joined Y Combinator’s Fall 2025 batch after initial development of their technology. Unsiloed differentiates itself through “dual-stream” vision models that process text and image data separately, alongside a technique known as “heat map-based chunking” that prioritizes key document areas such as figures and numbers. This approach also utilizes a specialized decoder that accounts for industry-specific terminology.
The company is currently serving clients in finance, legal, and healthcare sectors. One bank uses the technology for document processing, while an unnamed law firm has deployed it on its own servers for contract and case file management. Mishra also identified voice AI startup Bolna AI as a customer that uses Unsiloed to ingest knowledge bases for voice agents. Pricing starts at one cent per page, a cost reduction achieved by directing computing power only to high-value sections of documents.
In terms of performance, Unsiloed reported 99.78% accuracy across approximately 84,000 data points in a test involving 542 financial documents, outperforming major AI models and Microsoft’s Azure Document Intelligence, according to a technical brief shared by the startup. Additionally, the company claims to have topped the olmOCR-Bench public benchmark developed by Allen AI, scoring 88.0 compared to 66.0 for Reducto and 39.9 for Unstructured. These results have not been independently verified. Unsiloed plans to use the new funding to hire researchers and engineers, expand sales efforts, and develop proprietary datasets. The startup currently has a headcount of seven employees.
I use Bolna AI for my projects. Knowing they are now integrating Unsiloed for knowledge base ingestion is good news.
Seven employees with a $7.5M raise? They are going to have to hire a lot of engineers very quickly.
Dual-stream vision models processing text and images separately is a smart architectural choice for complex layouts.
Hallucinations in financial documents are unacceptable. Any source-tracing feature that allows human verification is a must-have.
Great to see another successful duo from IIT Kharagpur making waves in the US startup ecosystem this year.
One cent per page? That is surprisingly cheap compared to current market rates for structured data extraction services.
Law firms using on-premise servers makes sense for data privacy. Handling sensitive case files with cloud AI is still a big hurdle.
Those benchmark scores sound incredible, but who is verifying them independently? I usually need third-party validation before believing startup claims.
Felicis leading this round is impressive. Document processing is a massive bottleneck for enterprise AI adoption right now.