BREAKING
Vercel Resolves OpenTelemetry Blind Spot in AI SDK With New Inference Duration Histogram ● YBUSDT Approaches Potential Turning Point Amid Signs of Returning Buying Interest ● SEC to Resume Processing Registration Applications ● Rarible Unveils Rebuilt Cross-Chain Marketplace Featuring High-Speed Trading and New Rewards System ● Niko Bellic Appears in GTA 6: Voice Actor Michael Hollick Returns for Late-Game Cameo ● Ant International’s Alipay+ Connects Over 50 Global E-Wallets to Brazil’s Pix to Revolutionize Cross-Border Payments ● Federal Investigation Links Ransomware Negotiation Executives to ShinyHunters Hacking Group and FBI Data Breach ● White House National Cyber Director Sean Cairncross to Step Down Amid Transitioning Political Landscape and Unexplained Tech Outages ●
Venture Capital & Startups

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

Photo: Documentation / News Illustration

The fresh capital injection is designed to help the San Francisco-headquartered startup further develop its specialized technology, which aims to make unstructured enterprise documents far more accessible and digestible for AI agents. In addition to Felicis leading the financing round, the seed investment saw strong participation from Y Combinator, Orange Collective, and Uncorrelated Ventures. Prominent angel investors Aarthi Ramamurthy and Beth Turner also contributed to the round, signaling robust early-stage confidence in the startup’s vision and foundational technical architecture.

Modern business documents are rarely simple blocks of plain text. They frequently incorporate complex layouts, multi-column spreadsheets, detailed charts, and specialized visual elements that traditional and even many advanced AI systems struggle to interpret accurately. For critical sectors like banking, healthcare, and legal services, getting this information wrong can translate into exceptionally costly mistakes. Because these institutions rely heavily on accurate data extraction for their daily operations, compliance reporting, and client management, the potential operational and financial impact of an interpretation error can be severe.

Unsiloed aims to solve this fundamental enterprise pain point through sophisticated software designed to break down complex documents piece by piece. The platform extracts intricate information from tables, charts, and other visual elements, transforming it into clean, structured data that downstream AI agents can seamlessly consume and utilize. Furthermore, the technology incorporates a traceability feature that pinpoints exactly where the extracted information originated within the original source document. This allows enterprise users to independently check and verify the results rather than blindly taking the AI’s probabilistic output at face value.

The core challenge facing developers in the generative AI space stems from the underlying nature of large language models, which are inherently probabilistic rather than deterministic. Speaking in an interview, Unsiloed co-founder and CEO Aman Mishra explained that some degree of deviation and hallucination is practically inevitable when relying solely on traditional models, and that these minor discrepancies can create a massive negative impact further down the line in enterprise workflows.

Mishra co-founded Unsiloed alongside Adnan Abbas, a former classmate at the Indian Institute of Technology (IIT) Kharagpur. After spending crucial months developing and refining their core technology, the duo applied to and joined Y Combinator’s Fall 2025 batch, propelling their startup onto the radar of major venture capital investors.

Despite being a young startup, Unsiloed has already managed to secure early traction with enterprise customers operating in demanding sectors such as finance, legal services, and healthcare. Mishra shared that one major commercial bank is actively utilizing the startup’s technology to process incoming financial documents, while a law firm has deployed the software directly on its own private servers to handle sensitive contracts and archived case files. Additionally, Mishra named voice AI startup Bolna AI as a customer, noting that Bolna utilizes Unsiloed to ingest and structure extensive knowledge bases required for building responsive voice agents.

Entering the market for intelligent document processing places the young San Francisco startup into an already crowded and fiercely competitive arena. Unsiloed must contend with specialized document-processing startups such as Reducto and Unstructured, as well as heavily funded enterprise offerings provided by cloud giants like Microsoft and Amazon. Furthermore, businesses frequently attempt to leverage general-purpose AI models directly for document extraction tasks, giving Unsiloed yet another layer of competition to navigate as it seeks to carve out a distinct market share.

Addressing how the startup intends to stand out against this formidable backdrop, Mishra told TechCrunch that Unsiloed’s primary competitive advantage lies in its novel approach to processing complex documents. The startup utilizes what it terms "dual-stream" vision models, an architecture designed to handle text data and image data through completely separate processing streams. This is paired with a proprietary technique known as "heat map-based chunking," which intelligently identifies and prioritizes the most critical and information-dense parts of any given document.

Describing the technical implementation, Mishra noted that in one stream the system processes exact text tokens, while in the other stream it handles exact image tokens, an architecture that the founding team spent a significant amount of time designing and optimizing.

Once the document is parsed through this dual-stream framework, the technology dynamically directs heightened computing power toward areas containing complex figures, dense numbers, and intricate charts. Additionally, Unsiloed incorporates a specialized decoder that takes industry-specific terminology into account. This linguistic awareness helps the system accurately distinguish between terms that can carry drastically different meanings depending on the specific business or regulatory context in which they appear.

To back up its performance claims, Mishra pointed to a rigorous internal test involving 542 financial documents supplied directly by an enterprise customer. According to a technical brief shared by the startup, Unsiloed achieved an impressive 99.78% accuracy rate across approximately 84,000 individually extracted data points. The test compared Unsiloed’s performance against several major out-of-the-box AI models, Microsoft’s established Azure Document Intelligence platform, and the customer’s own proprietary in-house data extraction tool. However, it is worth noting that these specific test results have not yet been independently verified by third-party auditors.

In addition to the customer-specific financial document benchmark, Unsiloed claims to have significantly outperformed competing AI document-processing tools on olmOCR-Bench, a public evaluation benchmark developed by Allen AI. In benchmark tests conducted by the startup, Unsiloed achieved a score of 88.0, comparing favorably against Reducto’s agentic system, which scored 66.0, and Unstructured, which registered a score of 39.9. The benchmark evaluation covered 1,403 complex PDF pages and rigorously tested how effectively the various systems handled difficult formatting elements such as data tables, mathematical expressions, scanned legacy documents, and intricate multi-column layouts.

On the commercial front, Unsiloed’s pricing model starts at a competitive rate of one cent per page, according to Mishra. He asserted that the startup’s intelligent, resource-allocating approach also helps keep internal operational costs down by directing heavy computing power exclusively toward the sections of a document that genuinely require deep scrutiny, rather than brute-forcing the processing of every single page section uniformly.

Currently operating with a lean core team of just seven employees, the San Francisco startup plans to deploy its newly secured seed funding to scale its operations strategically. Mishra stated that the capital will be used primarily to hire additional top-tier research scientists and software engineers, aggressively expand the company’s enterprise sales and go-to-market efforts, and develop proprietary datasets designed to continuously improve and refine the underlying technology.

What is your reaction to this story?

Evan Lee Salim

Writer and contributor at Bitcoin 98.

Comments (0)