Healthcare AI · agentsKaggle Agents Intensive capstoneDec 2025
EmergencyCare Navigator
Multi-agent emergency triage that routes to the best facility and produces a clinician-ready SBAR handoff.
Problem
In a medical emergency people lose minutes to three questions: what matters right now, where to go, and what to tell the hospital so it can prepare. The goal was to cut that thinking-and-searching time while keeping triage conservative and rule-based.
What I built
- Four agents in sequence: IntakeAgent (short structured questions) → rule-based TriageAgent (safety-first escalation) → FacilityFinderAgent (Nominatim geocoding and facility search, OSRM ETA routing, distance-only fallback) → CoordinatorAgent (session state, SBAR handoff, approval flow).
- Emergency mode creates a pre-arrival hospital alert and pauses until the hospital acknowledges it; non-emergency mode generates a simulated appointment request. The same session continues with no context loss.
- session_id state plus a small JSON MemoryBank for minimal preferences; a trace_id per run, structured tool-call, error and retry logs, and metrics counters.
- Optional Gemini for calmer explanations; the triage decision never depends on the LLM. Six deterministic eval cases: chest pain, stroke signs, anaphylaxis, major bleeding, head trauma, mild fever.
- Not medical advice. Hospital alerts and bookings are simulated.
Stack
PythonPydanticNominatimOSRMGeminiMulti-agent