C Clinic AI

Top down view of our 3D hospital. With reception, ICU, general ward, discharge and staff room.
Dr. Patel making decisions during the live simulation.
Situation report generated at every tick. Shows general stats and a short overview paragraph.
Event log showing doctors choices and patients movement at every tick.
Different doctors with different levels of work and therefore different AI decisions.
End of simulation analysis report.
End report generated by AI at end of a simulation detailing considerations and next steps.

Inspiration

Healthcare systems around the world face constant resource pressure - too few doctors, too few beds, too many patients. We wanted to visualise what that actually looks like in real time, and ask: can we use AI agents to model hospital entities and their interactions? C Clinic was born from the question of whether LLM-backed agents could model the complexity and cascade effects of a hospital under stress.

What it does

C Clinic is a real-time hospital simulation where every patient and doctor is a live AI agent. Doctors use LLM decision-making and personas to make triage decisions - choosing who to treat next based on severity, wait time, and bed availability - and can explain their reasoning on demand. Users can watch the simulation run at normal pace, then trigger a Mass Casualty Surge or Staff Shortage with a single click and observe the impact: ICU fills up, queues back up, throughput drops. A live map shows agents moving between the Waiting Room, General Ward, and ICU, while charts track occupancy, queue depth, and throughput in real time. Every parameter - arrival rate, tick speed, doctor count, bed count - can be tuned live from the control panel without restarting. When the simulation is stopped, an LLM-written Session Report analyses the full run - identifying phases, key interventions, mortality rate, and strategic recommendations.

How we built it

We split the project into four parallel workstreams with clean interfaces between them:

Simulation Engine (Python + asyncio) - stateless tick loop managing patients, doctors, wards, and queues LLM Layer (Anthropic SDK) - injected as a callback into the engine; doctors call Claude Haiku for triage decisions and on-demand explanations Backend API (FastAPI + WebSockets) - broadcasts SimulationState every tick; exposes REST endpoints for scenarios and config Frontend (React 19 + TypeScript + Zustand + Recharts + Tailwind CSS v4) - live hospital map rendered with animated icons, three real-time charts, and an entity detail panel that shows AI summaries The LLM layer is decoupled via an LLMInterface protocol, so the engine runs and is fully testable with rule-based fallback logic when no API key is present.

Challenges we ran into

Keeping the simulation fast with async LLM calls - doctor decisions are non-blocking, but we had to be careful not to let queued LLM requests cause drift between ticks WebSocket fan-out at high tick rates - broadcasting full state every second to multiple clients required careful serialisation and connection management Coordinating four independent workstreams - agreeing on data contracts upfront (shared SimulationState types) was the only thing that made parallel development possible

Accomplishments

Every entity on the map is a genuine AI agent - not a scripted animation but a live decision-making loop. The explainability panel: clicking a doctor and reading its LLM-generated reasoning in real time is useful for analysis. At the same time, without an API key, it falls back to rule-based logic and still runs a full demo.

What we learned

LLM agents work best when you give them a tight, structured context window, as verbose prompts with irrelevant state produce worse decisions and higher latency. Decoupling via protocol interfaces (LLMInterface) was worth the upfront overhead; it let us test the engine independently and swap in the real LLM with zero changes to simulation logic.

What's next for C Clinic AI

Multi-hospital routing: ambulances should choose between hospitals based on real-time capacity. Real data integration would seed arrival rates and severity distributions from NHS or public hospital datasets for realistic scenario modelling. Additionally, live expansion of the hospital on a 3D model would allow hospitals to import their layout and model the simulation accurately in their hospital.