Real-time community threat intelligence. Users submit anonymous safety tips, a multi-agent AI pipeline verifies and escalates them, and everything is visualized on a 3D Mapbox map.
- Submit flares — click any building on the map and report a safety incident (active threat, weather, infrastructure, general safety)
- AI analysis pipeline — each tip is automatically processed by 4 sequential agents: Classifier → Corroborator → Synthesizer → Recommender
- Live receipt — every tip card in the feed shows a real-time AI analysis receipt that fills in as agents complete, streamed over WebSocket
- Escalation — when 3+ corroborated tips hit the same area, the building turns red with a pulsing alert and evacuation guidance is generated
- Community exits — users can contribute mapped emergency exits for buildings, which the Recommender agent uses when generating evacuation guidance
- Credibility scoring — tips carry a credibility score influenced by upvotes and AI classification
| Layer | Tech |
|---|---|
| Framework | Next.js 14 (App Router) |
| Map | Mapbox GL JS — dark-v11 + 3D building extrusions |
| Database | MongoDB (Mongoose) |
| Realtime | Socket.IO on a custom Node HTTP server |
| AI | Gemini 2.0 Flash Lite → FeatherlessAI fallback |
| Auth | JWT httpOnly cookie, bcrypt |
| Styling | Tailwind CSS + inline frosted-glass design system |
The pipeline runs async after each tip is submitted. Agents emit Socket.IO events so the UI updates in real time.
- Classifier — assigns threat level (
info/advisory/warning/critical), credibility adjustment, urgency decay, and source tags (firsthand,immediate, etc.) - Corroborator — cross-references nearby tips from the last 30 minutes, outputs a confidence score and escalation flag
- Synthesizer — only runs on escalation; produces a unified situation report and key facts
- Recommender — only runs on escalation; generates evacuation direction, shelter advice, and areas to avoid using mapped exit data
Primary model: gemini-2.0-flash-lite. Falls back to meta-llama/Meta-Llama-3.1-8B-Instruct via FeatherlessAI on quota errors.
- Node.js 18+
- MongoDB instance
- Mapbox account
- Gemini API key (Google AI Studio) and/or FeatherlessAI API key
npm installCreate .env.local:
NEXT_PUBLIC_MAPBOX_TOKEN=your_mapbox_token
GEMINI_API_KEY=your_gemini_key
FEATHERLESS_API_KEY=your_featherless_key # fallback if Gemini quota is hit
MONGODB_URI=mongodb://...
JWT_SECRET=your_jwt_secretnpm run devThe app runs on http://localhost:3000. The dev server is a custom Node.js server (server.js) that hosts both Next.js and Socket.IO on the same port.
| Command | Description |
|---|---|
npm run dev |
Start development server |
npm run build |
Build for production |
npm start |
Start production server |
src/
├── app/
│ ├── api/
│ │ ├── tips/ # CRUD, upvote, analyze (re-run pipeline)
│ │ ├── exits/ # Community-contributed emergency exits
│ │ ├── threats/ # Active threat buildings
│ │ ├── emergencies/ # Emergency feed
│ │ └── auth/ # Login, signup, session
│ └── dashboard/ # Main map view
├── components/
│ ├── Map.tsx # Mapbox map, 3D buildings, exit markers
│ ├── NotificationFeed.tsx
│ ├── AnalysisReceipt.tsx
│ ├── TipModal.tsx
│ ├── BuildingPopup.tsx
│ └── ExitModal.tsx
├── lib/
│ ├── agents/ # orchestrator, classifier, corroborator, synthesizer, recommender, llm
│ ├── db/ # MongoDB models: tips, users, exits, emergencies
│ └── socket.ts # Server-side Socket.IO emit helper
└── hooks/
├── useTipAnalysis.ts # Socket subscription + live agent state
└── useUserLocation.ts
| Variable | Required | Description |
|---|---|---|
NEXT_PUBLIC_MAPBOX_TOKEN |
Yes | Mapbox public token |
GEMINI_API_KEY |
Yes | Google Gemini API key |
FEATHERLESS_API_KEY |
No | FeatherlessAI key (Gemini fallback) |
MONGODB_URI |
Yes | MongoDB connection string |
JWT_SECRET |
Yes | Secret for signing auth tokens |