mydocumentintelligence.com — AI-powered legal document intelligence: clause extraction, risk scoring, and due diligence for legal teams who need to know where the answer came from.
The system reads structured and unstructured agreements — MSAs, NDAs, employment contracts, leases, financing documents — and returns a structured brief a reviewer can actually act on. The design constraint that shapes everything else: a lawyer cannot use an answer they cannot verify, so every output carries a citation back to a page and paragraph.
Retrieval is hybrid: BM25 for the exact tokens legal text is full of — party names, statute citations, section numbers, defined terms in quotes — fused with dense vectors for the paraphrase cases, combined by Reciprocal Rank Fusion. Chunking landed at roughly 800 tokens with 100-token overlap, with a hard rule never to split inside a numbered clause; the conventional 512/64 split fragmented clauses across boundaries and measurably hurt context recall.
Multi-tenancy is enforced twice — a tenant filter at the application layer and again as a Postgres row-level security policy — because a missing filter is a data-leak bug, not a quality bug. Serving is FastAPI on AWS Fargate behind CloudFront and WAF, with pgvector on RDS, documents in S3 under per-tenant prefixes with KMS encryption, and secrets retrieved at boot through the task IAM role.
There is a full engineering write-up of the architecture on this site: MyDocumentIntelligence — Architecture Deep Dive.