AI incident command copilot for smart stadiums and tournament operations.
Live demo: https://matchday-command-1069049902747.asia-south1.run.app
Built solo for Google PromptWars (theme: Smart Stadiums & Tournament Operations). While most stadium AI targets fans, MatchDay Command sits on the other side of the glass — in the operations control room. Live gate, crowd, radio, medical, and weather signals stream into one dashboard; Gemini triages incidents, recommends staff actions, and writes the operational briefings and shift-handover reports that ops teams produce by hand today.
Running 104 World Cup matches across 16 stadiums pushed real tournaments toward digital twins and predictive crowd management. The expensive part isn't the sensors — it's the human bottleneck: an ops room drowning in radio chatter and dashboards, making judgment calls under pressure. MatchDay Command demonstrates that a well-prompted LLM can carry the cognitive load: cluster the noise into incidents, explain the risk, propose the next action, and write it all down.
The app has three views on one shared live match: the ops control room (/), the staff console (/staff), and the tournament supervisor (/tournament). One server-side simulation broadcasts to every client — all screens show the same clock and story, and sim controls (speed / next-beat / restart) are global.
| Ops control room | Emergency evacuation mode |
|---|---|
 |
 |
The staff console (/staff) — a role-grouped task board fed by incidents, weather, traffic, and the AI demand plan:
- Live signal feed — a match-day simulator streams gate throughput, crowd density, steward radio logs, medical calls, and weather over SSE. (Simulated data means the demo never depends on hardware or third-party APIs.)
- AI incident triage — one click sends an incident plus its source telemetry to Gemini, which returns structured output: severity, rationale, prioritized actions assigned to radio channels, and an escalation flag.
- Ops briefings & handovers — Gemini drafts the periodic situation briefing and the end-of-shift handover report from the event log.
- Approach traffic & routing — a live Google Map with real-time traffic (when a Maps key is present; a schematic fallback otherwise) shows approach-corridor congestion, parking-lot fill, and AI-style routing advisories ("steer arrivals to Ring Road, direct parking to Lot C").
- Weather-driven demand forecasting — real forecast via Open-Meteo (keyless) feeds Gemini, which predicts supply demand (umbrellas, ponchos, water, cold drinks, hot food, fans, blankets) against on-hand stock and flags shortfalls.
- Staff dispatch console — the system doesn't just watch, it acts: prep/stocking orders, traffic advisories, weather readiness, and incident response are dispatched as role-grouped tasks with an ack → in-progress → done lifecycle.
- Control-room copilot (agentic) — a chat that answers from live state and takes actions via Gemini function calling: "tell catering to start hot food prep" dispatches a real task to the staff console; it can also open incidents and trigger briefings.
- Predictive early warnings — trend projection over live telemetry warns before thresholds break: "North Concourse on track to hit 85% in ~9 min" — with one-click pre-emptive incident creation.
- Voice radio reports — push-to-talk (browser Web Speech API, keyless): a spoken steward report lands in the live feed and flows into AI triage like any other signal.
- Tournament supervisor — every venue on one wall: the live stadium plus sister venues with staggered kickoffs, aggregate incident counts, and a highest-pressure ranking.
- Emergency mode — a guarded red button flips the venue to evacuation posture: gates go exit-only, Gemini writes the zone-by-zone evacuation plan and the PA announcement, and work orders are auto-dispatched to every staff role.
flowchart LR
SIM[Match-day simulator<br/>scripted scenario + noise] -->|SSE /api/stream| UI[Ops dashboard<br/>Next.js App Router]
SIM --> STORE[(In-memory OpsState)]
UI -->|POST /api/triage| GEMINI[Gemini Flash<br/>structured output]
UI -->|POST /api/briefing| GEMINI
STORE --> GEMINI
Everything shares one vocabulary: the types in src/shared/models/ define events, incidents, briefings, and the SSE wire protocol — used by the simulator, the API routes, and the UI alike. See docs/ARCHITECTURE.md for the design overview, and docs/agents/ (entrypoint: AGENTS.md) for the extensive documentation set — architecture, domain models, simulation, AI layer, frontend, full API reference, infrastructure, and conventions.
- Next.js 15 (App Router, TypeScript, Tailwind v4) — one full-stack app
- Gemini API (
@google/genai, structured JSON output) — triage, briefings, demand forecasting - Google Maps Platform (Maps JavaScript API + live TrafficLayer) — approach traffic, with a keyless SVG fallback
- Open-Meteo (free, keyless) — live weather, with a simulated fallback
- Server-Sent Events for the live feed
- Cloud Run for deployment (Docker, standalone output)
| Var | Purpose | Required |
|---|---|---|
GEMINI_API_KEY |
Triage, briefings, demand (Gemini) | Yes for AI features |
GEMINI_MODEL |
Primary model id (default gemini-3.5-flash) |
No |
GEMINI_FALLBACK_MODELS |
Comma-separated models tried if the primary is overloaded (503/500/429) | No |
MAPS_API_KEY |
Google Maps live traffic; delivered to the client at runtime via /api/config. Without it, the map degrades to a schematic. |
No |
Everything degrades gracefully: with no keys at all, the feed, map (schematic), and dashboards still work — only the Gemini-backed panels surface an error.
npm install
cp .env.example .env.local # add your Gemini API key (free at aistudio.google.com/apikey)
npm run dev # http://localhost:3000Useful checks: npm run lint, npm run format, npm run typecheck, npm test, npm run build.
Unit tests (Vitest) cover the deterministic core — the logic where correctness matters and there are no external dependencies:
-
Simulation & detection — incident detection rules, early-warning trend projection, match-phase folding
-
Deterministic models — approach-traffic/parking, tournament summaries, weather live/fallback mapping
-
AI layer — the Gemini model-fallback chain (mocked client) and transient-error classification
-
Client state — the SSE reducer's dedupe / reset / zone-gate projection logic
-
Dispatch store — id assignment, status defaults, queue cap
-
Routes & security — API request-body validation (400s) and per-client rate limiting (429s)
npm test # run once
npm run test:watch # watch mode
npm run test:coverage # coverage floor enforced (~77%)Accessibility is audited with Playwright + axe-core against the running app — static pages plus interactive states (copilot, modals) — and passes with zero WCAG 2.1 AA violations:
npm run dev # in one terminal
npm run audit:a11y # in anotherCI (.github/workflows/ci.yml) runs Prettier check, ESLint, typecheck, the full test suite (with coverage floor), and a production build on every push and pull request. Security posture is documented in SECURITY.md.
gcloud run deploy matchday-command --source . --region asia-south1 \
--allow-unauthenticated --max-instances 1 \
--set-env-vars GEMINI_API_KEY=<your-key>(--max-instances 1 because match-day state is in-memory by design — see src/lib/store.ts.)
- Shared domain models + SSE protocol
- Simulator with scripted match-day scenario + baseline arrival curves
- Gemini triage endpoint (structured output)
- Live ops dashboard (zone density meters, gate tiles, event ticker, incident queue)
- Incident detection rules (density/queue/medical/weather → auto-open incidents)
- Briefing + shift handover generation
- Cloud Run deployment (live demo)
- Judge-facing replay controls (deterministic "next beat" fast-forward) + report export
- Mobile layout (incidents above feed, stacked panels)
- Approach traffic map (Google Maps live traffic + schematic fallback)
- Weather-driven demand forecasting (Open-Meteo + Gemini)
- Staff dispatch console (
/staff) with role-grouped task board - Agentic copilot chat (Gemini function calling → open incidents, dispatch tasks, briefings)
- Predictive early warnings (trend projection ahead of threshold breaches)
- Push-to-talk voice radio reports (Web Speech API)
- Shared live match state (one server-side sim, all views on the same clock, global controls)
- Tournament supervisor view (
/tournament) - Emergency mode (AI evacuation plan + PA text + auto-dispatched staff orders)