Unsiloed AI reads the documents everyone else gives up on
A two-person startup out of Y Combinator turns PDFs, scans and spreadsheets into data machines can actually use. It also happens to top a public benchmark full of the biggest names in AI.
Pull up a real bank document and it rarely looks like the tidy example in a product demo. There is a table that runs onto a third page, a signature bleeding across two cells, a chart with no legend, and a footnote referring to a schedule that lives in a different file. A human analyst learns to read around all of it. Most software does not. That gap - between what a document contains and what a machine can actually use - is the entire business of Unsiloed AI.
The San Francisco company, founded in 2024 by Aman Mishra and Adnan Abbas, sells an API that converts unstructured documents into structured data. In practice that means you hand it a PDF, a scan, a spreadsheet or an image, and it hands back clean JSON and LLM-ready Markdown that an AI model or agent can reason over. It is not a chatbot and not a search box. It is the layer underneath both.
That layer turns out to be where a lot of AI projects quietly stall. Unsiloed cites a figure that anyone who has worked in enterprise data will recognize: roughly 80% of enterprise data is unstructured, locked in formats that were built for people to read, not machines to parse. The company also points to a second, harder number - fewer than 10% of in-house document pipelines ever reach production. Teams start them, spend months, and shelve them.
Source: Unsiloed AI. Figures describe the problem space the company targets.
The pitch is small enough to fit in one line
Unsiloed describes itself as an "API for parsing multimodal unstructured data." The product breaks into three verbs, and the simplicity is deliberate. Parse turns a document into structured Markdown while keeping the layout, tables and hierarchy intact. Extract pulls the specific fields you ask for into JSON, with typed values, confidence scores and bounding boxes that point back to where each value came from. Split takes a fat multi-document file and breaks it into retrievable chunks with parent-child indexing, which is what retrieval and agent pipelines need to work reliably.
Parse
PDFs, scans, images → LLM-ready Markdown, structure preserved
Extract
Typed JSON fields with confidence scores + bounding boxes
Split
Long files → retrievable chunks with parent-child indexing
The receipts: #1 on a public benchmark
Claims about accuracy are cheap in this corner of AI, so the more interesting move is that Unsiloed published its. In May 2026, the company's Unsiloed Parser v3.1 posted a strict pass rate of 88.0 on olmOCR-Bench, a public benchmark of 8,413 unit tests across 1,403 PDF pages. That put it first out of 19 services tested. On the semantic-correctness evaluation - an LLM-as-judge check of whether the meaning survives - it scored 94.8.
olmOCR-Bench · strict pass rate
The company did one more thing worth noting: it open-sourced the benchmark harness on its GitHub, so the run can be reproduced rather than taken on faith. For a product whose entire value is trust in an output, publishing the test rig is the argument. The margin over the second-place service was narrow - 88.0 to 87.4 - but the more telling comparison is the spread over the general-purpose models. A parser built by two people beat GPT-5.5 by more than three points and the older document tools by a wide margin, on a test where the data and scoring were open for anyone to inspect.
Benchmarks are a snapshot, and the leaderboard will move as everyone ships new versions. Unsiloed's own scores come from its v3.1 release, and the field it beat will not stand still. What the result buys the company is not a permanent crown but a credible answer to the first question any technical buyer asks: does this actually work on hard pages, and can I check?
Why it reads the hard pages
Under the API, Unsiloed uses what it calls dual-stream vision models. One stream reads content, the other reads layout and hierarchy, and the two run together instead of in sequence. The system first segments a page into text, tables, images and plots, then applies a domain-specific decoder trained on millions of enterprise documents. That decoder carries ontologies for finance, healthcare and legal work, so a field on a loan tape and a field on a lab report are not treated as the same anonymous string.
There is also a validation step that matters more than it sounds. The system checks cross-field constraints - that totals match their line items, that references resolve to something real. Anyone can extract a number. Extracting a number that is internally consistent with the rest of the document is the difference between a demo and a system a bank will run unattended.
Who is buying, and who is building
The customer list skews to places where a wrong number is expensive. Unsiloed says it processes millions of pages a week for Fortune 150 banks and NASDAQ-listed companies, and offers on-premise deployment for teams that cannot send documents to a third party. It also serves more than 10 fellow Y Combinator startups working in finance, legal and healthcare - the kind of early adopters who will tell you fast when an extraction is wrong.
The people building it come from the industries they now sell into. Aman Mishra, the CEO, is an IIT Kharagpur graduate who built ultra low-latency trading systems managing billions at a hedge fund, then worked as a founding engineer at a San Francisco AI copilot startup whose users included Goldman Sachs and Charles Schwab. Adnan Abbas, the CTO, studied at IIT Kharagpur and MIT, built multimodal models deployed at a Fortune 10 company, and worked on autonomous navigation at Mercedes-Benz R&D. Both have shipped things that cannot be wrong, which is a useful instinct to bring to document data.
That background shows up in how the product is aimed. Extraction outputs come with confidence scores and bounding boxes, which is what you need when a compliance team asks where a figure came from. On-premise deployment is offered because a bank or hospital often cannot send a document to an outside API at all. These are not features a consumer product would prioritize. They are the requirements of buyers who answer to auditors, and they are the reason a young company can win business from institutions that are usually the slowest to adopt anything new.
The business model, in plain terms
Unsiloed sells access to its API, the way most developer infrastructure does. A team wires Parse, Extract or Split into its own software and pays to run documents through them, rather than hiring engineers to build and maintain a parsing stack in-house. The pitch to a buyer is a straightforward trade: the six months and the specialized headcount a document pipeline usually costs, exchanged for three endpoints that already work. For regulated customers, the on-premise option keeps the data inside their own walls while still using Unsiloed's models.
The developer surface is small on purpose, and the company backs it with a public cookbook of recipes on GitHub - its most-starred repository. The strategy is familiar from other infrastructure companies: make the first call easy, let engineers try it on their own documents, and win on whether the output holds up. In a category where the buyer can test the product against their own worst PDF in an afternoon, being right on the hard page is the whole go-to-market.
Where it sits in a crowded market
Document parsing is not an empty field. Unsiloed competes with LlamaIndex's LlamaParse, Unstructured.io, Reducto, Extend, Nanonets and Datalab, and it competes indirectly with the general-purpose vision models from OpenAI, Anthropic, Google and Mistral that can also read a page. The benchmark result is the company's answer to the second group: a purpose-built parser beat the generalists at their own test.
The bet underneath the business is a quiet one. While most attention in AI goes to the model that talks back, Unsiloed pointed itself at the plumbing - the step that turns a document into something the model can even see. It is an unglamorous layer, and that is arguably the point. Plumbing that a bank depends on is hard to rip out.
For now the company is small - two founders and an early team, hiring for founding engineer and researcher roles, backed by Y Combinator with a seed round reported in late 2025. Whether it becomes a standard piece of enterprise AI infrastructure will be decided over years, not benchmarks. But the shape of the wager is clear: if AI agents are going to work inside real companies, someone has to make the documents readable first. Unsiloed AI is trying to be that someone.