Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableTechCrunch→ original

Unsiloed Raises $7.5M to Help AI Agents Read PDFs: 99.78% of Values Correct in Its Own Test

An agent reading a PDF report can mix up table columns, and a bank ends up with a wrong conclusion. Startup Unsiloed raised $7.5M in a seed round led by Felicis. It turns such files into data, with each value linked to its location in the original. In its own test on 542 financial documents, it extracted 99.78% of about 84 thousand values correctly, but there is no outside verification. Run the parser on your own files.

AI-processed from TechCrunch; edited by Hamidun News
Unsiloed Raises $7.5M to Help AI Agents Read PDFs: 99.78% of Values Correct in Its Own Test
Source: TechCrunch. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTechCrunch ↗
Editor-in-Chief
On olmOCR-Bench (as measured by Unsiloed)88.0 versus 66.0 for Reducto
Key metric #2Pricing starts at one cent per page; the team has seven people
Customersa bank, a law firm running the system on its own servers, Bolna AI

The problem is simple: companies want to entrust work to AI agents, but their knowledge sits not in a tidy database but in PDFs: contracts, financial reports, medical histories, scans of old case files. Imagine an agent reading an annual report and, because of a complex layout, picking up a figure from 2024 instead of 2025. It then confidently delivers a wrong conclusion about a loan or a contract.

Startup Unsiloed AI is going after this gap. It raised $7.5M in a seed round (the first major venture round) led by Felicis.

The headline figure, 99.78%, is the share of correctly extracted values out of roughly 84 thousand fields in a test on 542 financial documents. The company ran the test itself, and there is no independent verification yet.

What Unsiloed Actually Does

The product works as a layer between the document and the agent. It parses the file, pulls data out of tables, charts and other elements, and converts it into a structured form the agent can work with downstream. One important detail: for every value, you can see where in the original it came from. A person can quickly check the result instead of taking the model's answer on faith.

Y Combinator, Orange Collective, Uncorrelated Ventures and angel investors Aarti Ramamurthy and Beth Turner also took part in the round. Founder and CEO Aman Mishra explains the need for the product this way: LLMs are probabilistic by nature, their answers will always contain some deviation and a share of hallucinations, and further down the chain this hits the final result hard. He founded the company with Adnan Abbas, a classmate from India's IIT Kharagpur. The team went through Y Combinator's fall 2025 program and now has seven people.

How It Works Under the Hood

The company calls its architecture "dual-stream." Think of one employee reading the text of a page while another looks at it as an image, and then they compare notes. One stream carries text tokens, the other image tokens. The second idea is "heatmap chunking": the system finds the most important areas of a page and spends more compute on figures, numbers and charts instead of processing the whole document uniformly. According to Mishra, this also cuts costs. There is also a dedicated decoder that accounts for industry terminology and helps distinguish words that mean different things in different fields.

Pricing starts at one cent per page. Customers include a bank, a law firm (which deployed the system on its own servers for contracts and old case files) and voice startup Bolna AI, which uses Unsiloed to load knowledge bases for voice agents. Mishra did not name the bank or the firm.

The Numbers and Their Caveats

There are two results. First: on 542 financial documents from an enterprise customer, Unsiloed reached 99.78% accuracy on roughly 84 thousand extracted values. It was compared against large AI models, Microsoft's Azure Document Intelligence and the customer's own in-house tool. Second: on the public olmOCR-Bench from Allen AI (1 403 PDF pages with tables, formulas, scans and complex layouts), the startup scored 88.0, Reducto's agentic system scored 66.0, and Unstructured scored 39.9. Unsiloed ran these measurements itself as well.

The market is crowded: there are Reducto and Unstructured, offerings from Microsoft and Amazon, and many companies simply send PDFs to a general-purpose model.

What to Do in Practice

The round is modest, but the signal is clear: document parsing quality is becoming a separate layer in the agent stack, and people are willing to pay for it. If you are building RAG or an agent for banking, legal or medical work, do not take anyone else's percentages on faith. Collect 100–200 of your own files with the nastiest tables and scans, label the ground-truth values and compare several parsers. Look not only at accuracy but also at whether you can verify the source of each number, and at whether you can install the system in-house if the data cannot go to the cloud.

Unsiloed will spend the money on hiring researchers and engineers, expanding sales and building its own datasets. In a year it will become clear whether a team of seven can hold its edge against giants with their clouds.

Story

Key Questions Answered

What Unsiloed Actually Does?

The product works as a layer between the document and the agent. It parses the file, pulls data out of tables, charts and other elements, and converts it into a structured form the agent can work with downstream. One important detail: for every value, you can see where in the original it came from. A person can quickly check the result instead of taking the model's answer on faith.

How It Works Under the Hood?

The company calls its architecture "dual-stream." Think of one employee reading the text of a page while another looks at it as an image, and then they compare notes. One stream carries text tokens, the other image tokens. The second idea is "heatmap chunking": the system finds the most important areas of a page and spends more compute on figures, numbers and charts instead of processing the whole document uniformly. According to Mishra, this also cuts costs.

What are the key details on The Numbers and Their Caveats?

There are two results. First: on 542 financial documents from an enterprise customer, Unsiloed reached 99.78% accuracy on roughly 84 thousand extracted values. It was compared against large AI models, Microsoft's Azure Document Intelligence and the customer's own in-house tool. Second: on the public olmOCR-Bench from Allen AI (1 403 PDF pages with tables, formulas, scans and complex layouts), the startup scored 88.0, Reducto's agentic system scored 66.0, and Unstructured scored 39.9.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…