works/

Polku2026

A phone line for suspected lung cancer. A patient calls and describes their symptoms, rules check them against published Nordic criteria, and a nurse gets one task with a time already proposed. When the nurse accepts, Polku calls the patient back and books it.

The language model only turns speech into typed symptoms. A deterministic engine does the matching and cites a source for every criterion. Red flags return "call 112" from the engine whatever the model says.

Before the caller is identified the agent can read nothing, and after that only their own record. The server enforces this and logs every tool call.

The model was Qwen3 on vLLM on a Verda GPU in the EU. We started with Gemma 3 12B and switched for native tool calls and faster turns.

Built over a weekend at Junction 2026 for AstraZeneca's challenge. Patients, Mobile ID and hospital calendars were synthetic, and the steps after CT were seeded. Archived after the event.

Stack
TypeScript
Next.js
PostgreSQL
vLLM
ElevenLabs
Junction

2026

Polku