AI Medical Review
Medical pre-triage based on the Manchester Triage System, from the patient's spoken description to the hospital queue.
Technical detail
The patient describes the symptoms by text or voice. The AI asks follow-up questions based on that description and assigns a risk color from the Manchester Triage System. Warning signs have fixed rules that take priority over the model, and every escalation is logged for audit.
The classifier uses function calling with two deterministic tools: vital sign ranges and room availability. The slot offered comes from the tool result, never from the model's text. The same clinical rules live in a single package, used by the backend and by an MCP server.
If the AI fails or hits its quota, a local rules engine keeps the flow running. The API key never reaches the browser.
Stack: React with Vite, NestJS, Drizzle with PostgreSQL, Qwen Cloud (text and speech transcription), MCP server.