Imported from tirthsuba15/SentinelActual (AGENTS.md). Install upstream with npx skills add tirthsuba15/SentinelActual. Copyright stays with the author.

SENTINEL — AGENTS.md

What This App Does

SENTINEL is a real-time multi-hazard disaster intelligence platform for incident commanders. Three live pipelines — wildfire ember transport prediction, seismic damage mapping, and AI-powered resource coordination — fused into a single Palantir Foundry Ontology. AIP agent reasons over all three simultaneously and speaks ranked recommendations via ElevenLabs.

Stack

• Frontend: React + Vite + Leaflet.js + WebSocket client → deployed Vercel
• Backend: Python FastAPI → WebSocket server + pipeline HTTP endpoints
• Database: SQLite (crew/shelter/hospital state, event log, session)
• Agent: Palantir Foundry Ontology + AIP Agent Studio + AIP Automate
• GPU: AMD Instinct MI300X, PyTorch ROCm (seismic CNN)
• Voice: ElevenLabs Turbo v2.5 streaming TTS
• APIs: NASA FIRMS (fire), Open-Meteo (wind, no key), USGS GeoJSON (seismic, no key)
• Design: Google Stitch → DESIGN.md → Claude Code (P3 owns this file)

File Structure

sentinel/ ├── DESIGN.md ← P3 generates from Google Stitch, lives at root ├── CLAUDE.md ← P3 creates, wires DESIGN.md into every CC session ├── AGENTS.md ├── PROGRESS.md / DECISIONS.md / BLOCKERS.md / ENV.md / GITHUB.md ├── frontend/ │ ├── src/ │ │ ├── App.jsx │ │ ├── components/ │ │ │ ├── MapPanel.jsx ← Leaflet map + 5 layer toggles │ │ │ ├── ActionQueue.jsx ← AIP recommendation cards + ElevenLabs trigger │ │ │ ├── EventFeed.jsx ← Chronological event log │ │ │ ├── TopBar.jsx ← 6 live counters + 3 sync dots │ │ │ └── LayerControls.jsx ← Toggle pills + mute + SIMULATE button │ │ ├── hooks/useWebSocket.js ← WS connection to backend │ │ └── constants/northridge.js ← Pre-computed Northridge 1994 damage grid │ └── vite.config.js ├── backend/ │ ├── main.py ← FastAPI app + WebSocket server │ ├── pipelines/ │ │ ├── wildfire.py ← FIRMS poll + ember particle sim (NumPy) │ │ ├── seismic.py ← USGS poll + CNN + GMPE + liquefaction │ │ └── coordination.py ← Foundry Ontology writes │ ├── models/seismic_cnn.py ← PyTorch CNN (MI300X/ROCm) │ ├── ai/ │ │ ├── prompt.py ← AIP agent prompt templates │ │ └── elevenlabs.py ← TTS synthesis + streaming │ └── db/schema.sql └── data/liquefaction.shp ← USGS CA liquefaction susceptibility

SQLite Tables

• seismic_events (id, magnitude, depth, lat, lng, detected_at, processed_at)
• damage_zones (id, event_id, cell_id, lat, lng, damage_prob, soil_type, created_at)
• fire_hotspots (id, lat, lng, frp, confidence, detected_at)
• ember_risk_zones (id, hotspot_id, lat, lng, probability, forecast_at)
• suppression_crews (id, crew_id, lat, lng, status, capacity, assigned_zone_id, updated_at)
• shelters (id, name, lat, lng, occupancy, capacity, damage_zone_id)
• hospitals (id, name, lat, lng, capacity, alert_level, damage_zone_id)
• evacuation_routes (id, origin_zone_id, dest_zone_id, status)
• action_cards (id, action_type, resource_id, zone_id, confidence, rationale, status, created_at)
• sessions (id, started_at, mute_state)

API Base URL

• Dev: http://localhost:8000
• Prod: Vercel (frontend) + Railway (backend)

Auth

None. Single-session commander interface.

Branch Conventions

• main → P1 only, never push directly
• p1/scaffold → P1 (infra + merges)
• p2/backend → P2 (FastAPI + SQLite)
• p3/ui → P3 (React UI + design system)
• p4/ai → P4 (pipelines + AIP + ElevenLabs)

Agent Rules

• All Claude Code sessions read this file automatically
• Claude Code handles ALL branching, committing, pushing
• Commit format: [P1|P2|P3|P4] phase-N: description
• Agent teams: 2-5 agents max, each owns specific files, no teams after 4pm
• P3: every Claude Code prompt must begin "Read DESIGN.md and CLAUDE.md first."