Unsiloed AI (YC F25) ads

Today we are launching Higher-Order Extraction. Most extraction systems answer one question: where is the number on the page? But the numbers that matter are often not on the page. They have to be computed from the ones that are. Something like f(g(h(x1, x2, x3))), where x1, x2, x3 are values on the page. Today, the way most AI teams get these answers is to extract the raw fields and then write bespoke computation pipelines on top, one for each document type they care about. Higher-Order Extraction removes that layer. We decompose the request into a directed acyclic graph (DAG). Leaf nodes are values extracted from the document. Compute nodes are values calculated from other nodes. We topologically sort the graph so nothing runs before its inputs are ready. For each compute node, a model writes a deterministic function. A second model reviews each function against the calculation it is supposed to perform. The full graph compiles into a single executable script. The model handles interpretation and the code handles computation. When the document format changes, we re-extract the leaves and regenerate the graph. No hand-maintained pipeline per template. Live in beta today at unsiloed [dot] ai.

n8n doesn't perform OCR itself. Its "Extract from PDF" node relies on the PDF's embedded text layer, so if that layer is missing or corrupted, the extracted text can be confidently wrong. Since there's no indication that this happened, bad data can flow through the rest of your automation unnoticed. We built an Unsiloed node for n8n. Drop it into your document extraction workflow and it reads what's actually on the page with clean Markdown and tables intact. Try it yourself: https://lnkd.in/dEd9gDpJ

Turn the documents like invoices, statements, reports and contracts from your Snowflake stage into queryable rows with Unsiloed. Use our Snowflake integration to extract structured data with typed fields, confidence scores and source citation, connect it to a Cortex Agent and query them in plain English. Try it on your own documents. Link in comments

Have you got documents sitting in a Snowflake stage? Invoices, statements, reports, contracts. You can turn them into queryable rows from inside the Snowflake UI. We built an integration for @Snowflake. A Python UDF sends files straight from a Snowflake stage to Unsiloed's API and returns typed columns with a confidence score and a source citation on every value. Hook it up to a Cortex Agent and query your documents in plain language. The whole setup is one worksheet, and your API key stays in a Snowflake-managed secret. Try it on your own documents. Set it up: https://lnkd.in/dX8wh-Dr Ready-to-run SQL: https://lnkd.in/d4vZjMTZ Unsiloed is more accurate than Snowflake's native extraction on hard documents and can help reduce your Snowflake bill: https://lnkd.in/dbCt_4Qc

Hermes Agent just got way better at documents. We built a Hermes Plugin that turns messy PDFs, tables, charts and scanned docs into clean Markdown + structured JSON the agent can actually use. → Upload any document → Unsiloed parses it with proper structure and tables → Hermes automatically adds it to the wiki + extracts key facts into memory → Search and recall across all your documents in future sessions Built with Nous Research + Unsiloed AI (YC F25) Completely free and open source GitHub link in comments

We built an open-source OCR + document processing CLI agent. Claude Code, Codex and other coding agents struggle with parsing really complex PDFs. So we built a CLI agent that can: 1. extract and parse merged cells, tables that continue across pages, and footnotes attached to rows 2. everything comes back as structured JSON with confidence scores for each field 3. if you give it a file with multiple documents inside, it classifies and splits them In the video we give it NVIDIA's 10-K and ask for data that's spread across pages. Within a minute it came back with exact numbers, confidence scores, and citations to the source pages. Built with Unsiloed AI (YC F25) + LangChain Repo: https://lnkd.in/gGPVdukU