Sentinel Climate

Decision support for climate emergencies. Sentinel Climate watches public data feeds for climate disruption — extreme heat, flash floods, grid stress, dirty air, earthquakes, wildfires, tropical systems, hospital strain — scores the risk to any US county, and drafts a resource-allocation plan against that county's real inventory: cooling centers, ambulances, high-water vehicles, medical teams. The mission is timing: get the right resources moving toward the right place before the disaster peaks, not after.

It works for every US county (~3,144, from Census + ACS data), projects each county's risk 72 hours forward from the hourly forecast, and ranks a national risk board so you can see where the country needs resources most right now.

It is decision support, not decision making. Every assessment ends with a recommendation that a human emergency manager must review before anything deploys. That constraint is wired into the data model, not the marketing.

Live: https://sentinel-climate.vercel.app

How it works

  OSINT sources                     Multi-agent pipeline
 ┌──────────────┐                 ┌──────────────────────────┐
 │ NWS alerts   │                 │  SENTINEL  hazard watch  │
 │ USGS gauges  │   ┌─────────┐   │  MEDIC     health risk   │   ┌────────────┐
 │ OpenFEMA     │──▶│collector│──▶│  DISPATCHER logistics    │──▶│ allocation │
 │ EIA grid     │   │(snapshot│   │  COMMANDER decision      │   │   engine   │
 │ AirNow AQI   │   │ + prov- │   ├──────────────────────────┤   └─────┬──────┘
 │ CDC heat     │   │ enance) │   │ Claude (ANTHROPIC_API_   │         │
 └──────────────┘   └─────────┘   │ KEY) or deterministic    │         ▼
                                  │ rules fallback           │   ┌────────────┐
                                  └──────────────────────────┘   │ dashboard  │
                                                                 └────────────┘

1. Collector pulls a fresh snapshot of a county from ten public OSINT feeds and records, per source, whether the data was live, degraded, or unavailable.
2. Multi-agent pipeline — four agents in sequence: SENTINEL reads the hazards, MEDIC assesses population health risk, DISPATCHER works the logistics, COMMANDER makes the call (MONITOR / DEPLOY / EMERGENCY). With an ANTHROPIC_API_KEY the agents run on Claude; without one, a deterministic rules engine produces the same report shape and says so.
3. Allocation engine converts the risk score into resource demand, fills it from the county's inventory, and reports every shortfall with a mitigation — it never pretends inventory it doesn't have.
4. Risk trajectory projects the next 72 hours of risk from the NWS hourly forecast, so a county heading into a heat peak surfaces before it arrives.
5. National risk board sweeps the major metros (plus any watchlist counties) with the deterministic scorer and ranks them, turning "which county?" into a live leaderboard of where resources are most needed.
6. Dashboard renders the assessment with the provenance of every input, a county search over the full national registry, and the 72-hour sparkline.

Key endpoints

Endpoint	What
GET /api/counties/search?q=	Autocomplete over every US county
POST /api/agents/analyze {county}	Full multi-agent assessment
GET /api/allocation/plan/:county	Fast deterministic score + plan (no LLM)
GET /api/trajectory/:county	72-hour projected risk curve
GET /api/riskboard	National risk leaderboard
GET/POST/DELETE /api/watchlist	Counties to keep sweeping

Data sources

Source	What	Key required	Status when unconfigured
NWS (api.weather.gov)	Weather alerts, forecasts, hourly forecast, heat index	None	Live
USGS Water Services	River gauge height and discharge	None	Live
USGS Earthquakes (FDSN)	Recent seismic events near the county	None	Live
NIFC Wildfires	Active wildfire incidents near the county	None	Live
NHC Tropical	Active hurricanes / tropical storms	None	Live
OpenFEMA	Disaster declarations	None	Live
HHS Protect	State hospital inpatient / ICU occupancy	None	Live (degraded when the federal feed is stale)
Census + ACS	County registry, population, senior counts	None	Live (bundled)
EIA	Grid demand and reserve margin	EIA_API_KEY	Unavailable (labeled)
AirNow	Air quality index	AIRNOW_API_KEY	Unavailable (labeled)
CDC	Heat-health data	CDC_API_KEY	Unavailable (labeled)
Google CSE + OpenAI	Social listening / OSINT sweeps	GOOGLE_CSE_API_KEY, GOOGLE_CSE_ID, OPENAI_API_KEY	Unavailable (labeled)
ElevenLabs / Twilio	Voice and SMS outreach	ELEVENLABS_API_KEY, TWILIO_*	Demo mode (simulated, labeled)

All keys are optional and free to register. The app runs with zero keys — it just tells you, per source, that it is running degraded.

Quickstart

cp env.example .env    # fill in whichever keys you have; all are optional
npm install
npm run dev

Open http://localhost:5001. The server reads PORT from the environment and defaults to 5001 — port 5000 collides with AirPlay Receiver on macOS, so set PORT explicitly if you need a different one.

Checks:

npm run check   # TypeScript, strict mode
npm test        # vitest — allocation engine, all-hazards scoring, risk
                # trajectory, national county registry, risk board

Set ENABLE_SWEEP=1 (ignored on Vercel) to run a background sweep of the watchlist every SWEEP_INTERVAL_MINUTES (default 15) and collect alerts.

Data honesty

The core product rule: simulated data is never presented as live.

• Every assessment carries per-source provenance — a SourceStatus for each feed stating live, degraded, or unavailable, with a timestamp. The dashboard renders these as a provenance strip on every result.
• Every agent report is stamped with its mode: ai (Claude, model id included) or rules (deterministic fallback). The fallback never masquerades as AI.
• In rules mode, agent confidence is the fraction of data sources that were actually live — measured, not invented.
• Every decision output sets reviewRequired: true. A qualified human reviews before any resource moves. There is no auto-deploy path.

Origin

Built at the AI+ Expo hackathon in June 2025, now being hardened into a real decision-support tool: typed contracts shared between server and client, a deterministic fallback for every AI path, provenance on every payload, tests on the allocation math, national county coverage, all-hazards feeds, forward-looking risk projection, and a one-command deploy to Vercel serverless.