Health-impact simulator for data-center siting in Atlanta. Drop a pin on the map and see a forward projection of that neighborhood's health burden through 2030 under three scenarios: no new facility, a facility at the pin, and a facility at an auto-selected lower-burden alternative site within 10 miles.

client/   React + Vite + Tailwind + Mapbox GL + Recharts   (UI)
server/   Flask API + Prophet forecasting (Python venv)     (simulation + data layer)
tools/    Node: docx briefing generator + Playwright filer  (briefing + PSC submission)

For the full end-to-end system pipeline (data ingestion, Prophet ML, React UI, briefing, PSC automation), see PIPELINE.md.

The app runs fully offline. Live data sources (EJScreen, Census) fall back to bundled local data automatically, so missing network or keys never breaks the demo.

Run each block once, from the repo root.

cd server
python -m venv venv
venv\Scripts\python.exe -m pip install -r requirements.txt

cd ..\client
npm install

cd ..\tools
npm install
npx playwright install chromium

This generates the synthetic Atlanta fixtures and the merged GeoJSON the map + simulator read from. (The server also auto-builds the merged file on first boot if it's missing.)

cd ..\server
venv\Scripts\python.exe generate_data.py            # -> data/*.csv + epd_permits.json
venv\Scripts\python.exe -m app.services.ingest      # -> data/merged_data.json

VITE_MAPBOX_TOKEN=pk.your_token_here     # base map tiles (app data works without it)
VITE_API_BASE=http://127.0.0.1:5000      # Flask API base (default is fine)

Set one for a session in PowerShell with e.g. $env:CENSUS_API_KEY="..." before run.py.

Use two terminals.

# Terminal 1 — API  ->  http://127.0.0.1:5000
cd server
venv\Scripts\python.exe run.py

# Terminal 2 — UI   ->  http://127.0.0.1:5173
cd client
npm run dev

Open http://127.0.0.1:5173. The app preloads South Fulton so the demo works instantly. Drag the pin, search a neighborhood, download a briefing, or submit to the (dry-run) PSC record.

# API health
Invoke-WebRequest http://127.0.0.1:5000/health -UseBasicParsing

# A full simulation
Invoke-WebRequest http://127.0.0.1:5000/simulate -Method Post `
  -ContentType 'application/json' -Body '{"lat":33.631,"lng":-84.530}' -UseBasicParsing

# Inspect raw vs parsed live data sources (for tuning field mappings)
Invoke-WebRequest "http://127.0.0.1:5000/debug/sources?lat=33.631&lng=-84.530" -UseBasicParsing

/simulate orchestrates three sources, each with automatic local fallback:

• Every /simulate response includes a sources block reporting live vs local-fallback, surfaced in the right-panel footer.
• The burden time series is synthesized per tract by weighting the effective baseline score against cumulative permit-density growth year over year, then projected with Prophet.
• Bundled fixtures in server/data/ are realistic synthetic data from generate_data.py. Swap in real files with the same schema for production.
• Search is an offline local search over tract names (not Mapbox Geocoding).

/file writes into a real public record. It defaults to a dry run that drives the full Playwright fill flow against a local mock docket and returns confirmation text without touching the live PSC site. Live filing requires both confirm=true and FORESIGHT_ALLOW_LIVE_FILING=1, and the live selector mapping in tools/file_psc.js must be completed first. Do not enable live filing without authorization.