Unsiloed AI Raises $7.5 Million Seed and Claims 88.0 on olmOCR-Bench Against Reducto's 66.0
Startups / news
Unsiloed AI Raises $7.5 Million Seed and Claims 88.0 on olmOCR-Bench Against Reducto's 66.0
The Y Combinator Fall 2025 company reported the score itself, and its 99.78% accuracy figure on 542 financial documents has not been independently verified.

Unsiloed AI, a San Francisco startup from Y Combinator's Fall 2025 batch, raised a $7.5 million seed round led by Felicis, TechCrunch reported on November 18, 2026. The company sells software that turns PDFs into structured data for AI agents and says it scored 88.0 on a public OCR benchmark where Reducto scored 66.0.
Y Combinator, Orange Collective and Uncorrelated Ventures also invested, along with angels Aarthi Ramamurthy and Beth Turner. TechCrunch does not report a valuation or revenue. The round was announced; the article does not say when the cheque cleared.

Seven employees and a one-cent price
Co-founder and CEO Aman Mishra and co-founder Adnan Abbas were classmates at IIT Kharagpur. The company has seven employees. Pricing starts at one cent per page, so a 10,000-page archive costs $100 at the entry rate, though TechCrunch does not say what the higher tiers charge.
Customers are in finance, legal services and healthcare. An unnamed bank and an unnamed law firm, which runs the product on its own servers for contracts and old case files, are the largest references. The one named customer is Bolna AI, a voice-agent startup that uses it to ingest knowledge bases.
88.0 on olmOCR-Bench, from the company
The benchmark figure is company-reported. olmOCR-Bench is a dataset from Ai2, the Allen Institute for AI, of 1,403 PDF files and 7,010 unit tests, released under ODC-BY-1.0. TechCrunch describes it as 1,403 pages; the dataset card says files.
- Unsiloed AI88 points
- Reducto (agentic)66 points
- Unstructured39.9 points
Source: TechCrunch, Nov. 18, 2026, citing Unsiloed AI
The benchmark scores deterministic pass or fail checks: text presence, text absence, reading order, table accuracy and math formulas. Its composition matters for a startup selling to banks and law firms.
| Document source | PDFs | Tests |
|---|---|---|
| arXiv math | 522 | 2,927 |
| Tables | 188 | 1,020 |
| Multi-column | 231 | 884 |
| Headers and footers | 266 | 753 |
The dataset card lists three more sources, old scans, old scans with math, and long or tiny text, and does not explain how test results roll up into the overall score. Nearly half the PDFs are arXiv math papers, which look little like a credit agreement or a discharge summary.
A 99.78% figure that cannot be checked
Unsiloed also reported 99.78% accuracy on 542 financial documents holding about 84,000 data points, compared against major AI models, Microsoft's Azure Document Intelligence and a customer's in-house tool. TechCrunch notes the result has not been independently verified. At that rate roughly 185 of 84,000 points would be wrong, and the article does not say who chose the documents or how a data point was graded.
Mishra's design argument is that large language models are probabilistic and can introduce "deviation" and "hallucination". The product uses what he calls dual-stream vision models that process text and images separately, plus a heat map-based chunking step and a decoder tuned to industry terms. Each extracted item is traced back to its place in the source document so customers can verify it.
Reducto and Unstructured are named as competitors, alongside Microsoft and Amazon document offerings. For another seed-stage company in the agent-tooling business, see Maritime's $1-a-month agent VM, and for an Indian-founded agent-safety round, Disseqt AI's $8 million.
The money goes to researchers, sales and proprietary datasets. A customer naming itself, or an independent run of the 542-document test, would settle the accuracy claim.
Sources
More in Startups
- 01Onlook, a YC W25 Startup, Launches a Free Desktop Design Editor With 27,400 Stars and No Pricing PageThe Apache 2.0 app edits a team's own React code on a canvas through the user's Claude Code or Codex account, and neither source says whether the company has raised money.
- 02Genesis AI Raises $500 Million at $3.5 Billion With 30 to 40 Humanoid Robots BuiltThe two reports of the November 17 round disagree on the founding year and on Théophile Gervet's title, and neither discloses revenue or a pre- or post-money basis.
- 03Razorpay's Reported $5 Billion IPO Value Is 12 Times FY25 RevenueThe payments company is expected to publish its updated draft prospectus next week, which will be the first time anyone outside Sebi sees recent numbers.
- 04Maritime, a YC F26 Startup, Prices an Isolated Agent VM at $1 a MonthThe Launch HN post leads with the $1 figure; the pricing page it links to starts plans at $20 and caps how many machines stay awake.