Dwell

AI agent that turns scattered public government data into a city's early warning system, flagging which locations are at risk of infrastructure failure, why, and what to fix first, before residents feel it.

Demo video: https://www.youtube.com/watch?v=jGVV2ODN8nI

Dwell takes unstructured, disparate public data, soil reports, seismic records, weather and disaster history, building permits, complaint logs, and turns it into a structured infrastructure risk assessment: which hazards are active, how severe, on what general timeframe, and what to do about it. The goal is to surface infrastructure risk before it becomes failure, not by predicting exact failure dates, but by giving cities the structured, evidence based read they need to act early.

What it does

Give Dwell an address and it pulls real data from federal and municipal sources (soil and terrain, earthquake history, weather and FEMA disaster declarations, building permits and 311 complaints), scores the location across four hazard categories using a transparent, auditable rubric, grounds its recommendations in curated public works cost data, and produces:

• A plain language welfare summary of overall risk
• Four color coded hazard scores (green, orange, red): seismic, flood and ground, structural, infrastructure decay
• A prioritized, costed roadmap of concrete preventative actions
• Renter specific actions for people who do not own the property
• Honest disclosure of any data gaps encountered during the assessment

How it works

1. Gather. A Claude powered ReAct agent calls four data tools as needed: USDA SSURGO soil data, USGS earthquake history, NWS weather alerts and FEMA disaster declarations, and DataSF building permits and 311 complaints.
2. Score. A deterministic, hand built scoring rubric, not a trained model, turns that raw data into the four hazard grades. Every point is traceable to a specific data field, so a red grade always has a concrete, explainable reason.
3. Ground. A lightweight retrieval system pulls relevant public works knowledge (typical retrofit costs, pipe replacement ranges) so the roadmap's cost estimates are grounded rather than invented.
4. Report. The agent writes the final structured report using the rubric's grades as authoritative and the retrieved knowledge as its cost basis.

We deliberately did not train a predictive ML model. The available public data is current state snapshots, not labeled historical failure events, so a trained model would project false confidence it has not earned. An auditable rubric paired with an LLM that explains it is the more honest design, and it is what is running here.

Using the app

The frontend takes a street address, not a city or neighborhood name. Type a full address, for example 1 Dr Carlton B Goodlett Pl, San Francisco, CA, and click Analyze. The app geocodes the address, pulls the data, and generates the report.

Click + new in the sidebar to start a fresh report for a different address. Previous reports stay listed in the sidebar so you can revisit them without regenerating.

San Francisco only, for now

Soil, earthquake, and weather and FEMA data are national and will return results almost anywhere in the United States. Building permits and 311 complaints are currently scoped to San Francisco only, since that is the dataset we built and tested against for this hackathon. Addresses outside San Francisco will still generate a report, but the infrastructure decay and structural hazard scores will be based on partial data, and the report will say so explicitly in its coverage notes rather than presenting a falsely confident score. Adding another city means adding its open data sources to the building and infrastructure data module.

Setup

pip install -r requirements.txt
npm install
export ANTHROPIC_API_KEY=sk-ant-your-key-here

Run the backend and frontend in two terminals:

# Terminal 1
python3 api_server.py

# Terminal 2
npm run dev

Open the local URL Vite prints, typically http://127.0.0.1:5173.

Demo mode

Set DATAGENT_MODE=demo before starting api_server.py to serve cached responses for addresses you have already run once, instead of hitting live government APIs and the geocoder every time. Use DATAGENT_MODE=record once per address beforehand to populate that cache.

export DATAGENT_MODE=record
python3 api_server.py
# run your demo addresses once each, then:

export DATAGENT_MODE=demo
python3 api_server.py
# now those exact addresses load fast and without network dependency

Built with

Python, Anthropic Claude API, USGS Earthquake Catalog API, National Weather Service NWS Alerts API, FEMA OpenFEMA Disaster Declarations API, USDA SSURGO Soil Data Access API, DataSF Socrata Open Data API, FCC Census Block API, Census Geocoder API, React, Vite