Representative engagement — details anonymised.
Due Diligence Engine
Due diligence is thousands of pages read under deadline pressure. We built a system that reads all of them, cross-references what it finds, and never cites a page it can’t show you.

Challenge
An advisory team was spending the first two weeks of every engagement extracting the same facts from every data room: change-of-control clauses, indemnities, related-party transactions, covenant terms. Manual, repetitive, and error-prone at exactly the moment errors are most expensive.
Off-the-shelf AI tools summarised confidently and cited nothing. For work that ends up in front of a deal committee, an unverifiable answer is worse than no answer.
Approach
We treated hallucination as a structural problem, not a prompt problem. Every extraction runs through a typed pipeline: classify the document, segment it, extract against a schema, then verify each claim against the exact source span before it is allowed into the findings register.
Findings carry their evidence with them — page, paragraph, and confidence. Anything below threshold routes to a human review queue instead of the report.
The reviewing analyst stays in charge: the system drafts the findings register, the human signs it.
Outcome
First-pass document review that took two weeks now takes two days, and the partners trust it precisely because it refuses to answer what it cannot prove. The firm’s review capacity scaled without a single new hire.
- Claude
- TypeScript
- Postgres
- pgvector
- AI Pipelines
- Document Intelligence
- Compliance