spot_imgspot_img

Top 5 This Week

spot_img

Related Posts

IIT Kharagpur batchmates’ Unsiloed AI raises $7.5 million led by Felicis


Unsiloed AI, a San Francisco-based startup founded by IIT Kharagpur batchmates Aman Mishra and Adnan Abbas, has raised $7.5 million in a seed round led by Felicis. The startup builds software that turns complex documents into structured data that AI agents can use.

Advertisment

Y Combinator, Orange Collective and Uncorrelated Ventures also took part in the round, along with angel investors Aarthi Ramamurthy and Beth Turner. The startup did not disclose its valuation.

Unsiloed said it will use the funding to advance its document models, grow its engineering team and expand cloud and on-premises deployments for enterprise customers. Mishra, who is the startup’s CEO, also said the funds will go towards hiring researchers, building a larger sales effort and creating proprietary datasets to improve the technology.

What Unsiloed does

Unsiloed builds APIs (application programming interfaces) that let software read, extract and update information in documents such as PDFs, presentations, Word files, scanned forms, tables, charts and images.

The output is structured Markdown or JSON that large language models and AI agents can work with. The startup said its goal is to make documents “as computable as databases”.

Its platform handles document parsing, extraction, classification, splitting and editing, and is used across financial services, insurance, healthcare, legal and other industries.

The startup also cites where in the original document each piece of extracted data came from, so customers can check and audit the results. Pricing starts at one cent per page.

It supports private-cloud, on-premises and fully air-gapped deployments, which lets organisations process sensitive documents within their own infrastructure.

“We believe AI agents will eventually read and write more documents than humans,” Mishra said. “Those agents do not just need better OCR. They need a dynamic interface to read, update and act on unstructured data. We are building that interface so documents can become as computable as databases.”

Explaining why general-purpose AI models fall short on such work, Mishra told TechCrunch: “The problem with LLMs is they’re inherently probabilistic. There is going to be some deviation induced.”

How the technology works

Unsiloed uses what it calls a dual-stream architecture, in which vision models process a document’s content and its visual layout separately.

It also uses heat map-based chunking to detect pivot elements such as numbers and merged cells, keeping related information together, along with industry-specific decoders that account for differences in terminology across sectors.

“A document’s meaning depends on both its content and its visual structure,” said Adnan Abbas, co-founder and CTO. “We combine deterministic computer vision with dual-stream vision models to preserve those relationships, so AI agents can work with the information in context.”

The startup has made several performance claims based on its own testing. In a test involving 542 financial documents from an enterprise customer, it reported 99.78% accuracy across approximately 84,000 extracted data points, with its performance benchmarked against major AI models, Microsoft’s Azure Document Intelligence and the customer’s in-house tool.

It also scored 88.0 on olmOCR-Bench, a public document-reading benchmark from the Allen Institute for AI, in a May 2026 evaluation across 1,403 PDF pages.

By the startup’s numbers, Reducto’s agentic system scored 66.0 and Unstructured scored 39.9. Abbas said the system ranked first among 19 systems on the benchmark.

Customers and team

Unsiloed said it has processed millions of pages for organisations ranging from early-stage AI startups to publicly traded companies and law firms.

Mishra said one bank uses the technology to process documents, while a law firm runs it on its own servers for contracts and archived case files. Bolna AI, a voice AI startup, uses Unsiloed to load knowledge bases into its voice agents.

In its Y Combinator launch post, the company said it was processing millions of pages a week for Fortune 150 banks and Nasdaq-listed companies, and counted more than 10 Y Combinator startups among its clients.

The post also claimed that more than 80% of enterprise data is multimodal and unstructured, and that AI teams spend six months or more building accurate document-ingestion systems.

Founded in 2025, Unsiloed was part of Y Combinator’s Fall 2025 batch, according to YC. The startup has seven employees, while its founding team includes alumni of Cognition, Snowflake, Stripe and quantitative trading firms.

Before Unsiloed, Mishra built ultra low-latency trading systems at a hedge fund and was the founding engineer at a San Francisco startup building AI copilots for financial firms.

Abbas, who studied at MIT and IIT Kharagpur, built multimodal models deployed at a Fortune 10 company and worked on autonomous navigation systems at Mercedes-Benz.

“For AI agents to transform how enterprises work, they first need to understand the documents those businesses run on,” said Astasia Myers, senior venture partner at Felicis.

“Unsiloed is building the interface between those documents and AI agents, combining advanced computer vision algorithms and VLMs along with the deployment flexibility that enterprises need,” she said.

Unsiloed competes with large cloud providers such as Microsoft and Amazon, with general-purpose AI models that businesses sometimes use to pull information out of documents, and with specialist document AI startups.

One of those specialists is Reducto, also a Y Combinator alumnus based in San Francisco. Reducto raised $75 million in a Series B round led by Andreessen Horowitz in October 2025, taking its total funding to $108 million. At the time, it said it had processed nearly one billion pages for customers including Harvey, Rogo and Scale AI.



Source link

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles