Text in, triage out: 14 conditions, runs fully offline — no API keys, no cloud. 🎯 0.894 held-out accuracy, with 0.524 cross-source transfer recall published beside it 🔍 Every answer names the engine that decided and shows that engine's own accuracy 📄 616 rows, each quoting a CDC/WHO passage the builder checks verbatim 🗂️ SQLite consultations, audit trail, CSV export, Markdown reports, batch mode ✅ 117 tests, 11 documented endpoints I build NLP services whose numbers survive your technical reviewer. Triage demonstrator, not a clinical tool.