Inspiration
Every trainee call handler sits the same exam, but real emergency calls change depending on what you say. We wanted training that reacts to the trainee, not a fixed script.
What it does
NERVE is an adaptive simulator for training emergency call handlers. You take a call from an AI caller, who is voiced in real time. After every response, NERVE scores what you did and changes the call: it adds complications when you’re coping, cues you when you slip, and calms the caller when you’re overloaded. There are 20 scenarios, and each call ends with a debrief built from the recorded session.
How we built it
Frontend: React, TypeScript and Tailwind, styled after our launch film. Backend: FastAPI running a deterministic state engine and evaluator, which together decide the caller’s next move. Caller: a local Qwen model (via Ollama) writes each caller line under the engine’s instructions. A validator checks every line, with scripted fallback lines if one fails. Voice: ElevenLabs speaks the caller’s lines. Input: push-to-talk voice or typing.
Challenges we ran into
Keeping the AI caller consistent: it mustn’t reveal facts early, invent details or change the patient’s condition. Making 20 scenarios behave genuinely differently on one shared engine. Keeping the latency of a local model playable.
Accomplishments that we're proud of
A closed adaptive loop that works end to end with live AI and voice. 20 playable scenarios, each with its own call structure. 313 automated tests. A clean, cinematic interface. Clear labelling: one protocol demo awaiting clinical review, 19 scenarios that assess communication only.
What we learned
Let the engine own the plot and the AI write the scene. Deterministic rules plus a validated LLM beat a free-running chatbot for reliability, fairness and safety.
What's next for NERVE_simulator
Clinical validation of the scenarios with professionals. Presage camera-based physiology as extra context. It could only ever slow the escalation down, never make it harder. Time-series analytics with Tiger Data, and cloud deployment. New scenario packs for nurses, cybersecurity incident response and crisis teams.
