- Challenge
- A dual trust problem: fully private (no cloud or telemetry) and zero invented or missed clauses across many document types. Cloud AI is unsuitable for confidential legal review.
- Approach
- We shipped a Windows desktop app with guided legal task selection, production prompt packs, multi-format ingest, local LLM via Ollama (Phi-4 Mini class), streamed generation, PDF/Word/text exports, and a packaged installer. Scan-then-extract pipeline (regex clause scan to schema-scoped model call) instead of naive chunking.
- Result
- Zero cloud calls and zero telemetry. 12 to 15+ document types. Validated with real solicitors on real documents.
Related capability: TrustDesk AI for citation-first litigation intelligence in firm workflows.