suhaasborra.com

genx · agent runtime · hybrid rag
directory

blogdocumint-ai

documint ai

Legal documents are PDFs pretending to be structured data.

github profileblog post

the idea

the problem

A contract is a scan of a layout that carries meaning: clauses, parties, dates, obligations. OCR hands you a wall of text and drops the structure. Downstream tools then work out again what the page already knew.

the approach

Documint runs a six-stage pipeline. Every stage does one job — OCR, layout parsing, entity extraction, normalisation, relation building, knowledge graph — and passes a typed artefact to the next. The output is a graph of the document, not a transcript of it.

what's done

  • Six-stage pipeline from raw scan to knowledge graph.

  • Layout-aware parsing that keeps clauses attached to their headings.

  • Entity and relation extraction tuned for contract language.

  • A query surface on top of the resulting graph.

what's missing

  • Coverage outside contracts: filings, correspondence, handwritten annexes.

  • A hosted API. Today it runs as an internal service.

  • A public sample set to point people at.

how to try it

Private for now. Mail [email protected] if you have a document set worth testing.

Other public work: github.com/Suhaas-code.

all posts