AI review for government tender documents
A review workspace for procurement and engineering tenders: it parses the document, reviews it clause by clause with dual-model cross-checking against a versioned law library, and produces a tracked-change revision — with a person approving every finding.
- Long context
- A real 105k-character sample showed a single full-context pass recalls only 24 of 42 findings. I redesigned the review as structure-aware chunked deep scans with a fact ledger, fixed coverage matrices, and deterministic merging — 75% more findings, every one anchored to source text.
- Recall & cost
- Rebuilt clause retrieval so exact legal clauses reach the context 24/24 (previously widespread misses); reordered prompts for KV-cache prefix reuse — cache misses down 43.8%, cost per review down ~20% with zero lost findings. Frozen eval: Recall@3 90%, 100% abstention on negatives.
- Security
- AES-256-GCM envelope encryption, SQLCipher at rest, non-exportable device keys, Cloudflare Access; blue-green Docker deploys on a VPS.
- Quality
- 1,328 backend tests across 167 files, five CI workflows, fail-closed formatting gates on document output.
- 01 Upload and preserve originals
- 02 Parse and confirm metadata
- 03 Review against cited evidence
- 04 Accept, reject, or rewrite
- 05 Export a tracked revision
