FieldLine

Inspiration

After Hurricane Helene hit western North Carolina in 2024, a lot of help came from volunteers and local groups. Requests for help were posted on Facebook, sent by text, called into hotlines, and heard on radio, and there was no easy way for volunteers to see them in one place or know which roads were open. We wanted to build a tool that collects those requests and matches them to volunteers who can actually get there.

What it does

FieldLine takes in reports from social media, NGOs and shelters, SMS, public-safety radio, NCDOT road closures, and USGS river gauges. An LLM reads each report, pulls out what's needed and who's vulnerable, groups reports about the same household into one incident, and proposes a task.

For each incident, the system picks the best volunteer based on skills, vehicle, and drive time with closed roads removed. The coordinator can see why each nearby volunteer was picked or ruled out. The volunteer accepts the task in the phone app, drives there, and marks it complete once their GPS shows they're at the address. For the demo, we replay the first few days of Helene in Asheville.

How we built it

Backend: FastAPI, Tiger Cloud (TimescaleDB + PostGIS), OpenAI API
Frontends: React and MapLibre for the command dashboard and the volunteer app
Matching: PostGIS nearest-neighbor search, then routing on the OpenStreetMap road network with closed roads removed
Priority: AI-scored urgency, adjusted for the person's vulnerability and how hazardous their location is. The hazard comes from public rainfall, terrain, and CDC social vulnerability data.

Challenges we ran into

Geolocating vague reports. Most reports say things like "near the Dollar General on Leicester Hwy" instead of an address. We had the LLM pull out the road, landmark, and town, then matched those against the OpenStreetMap road network.
Sourcing relevant datasets. There's no single Helene dataset, so we pieced one together from NCDOT, USGS, NOAA, the Census, and the CDC, each in a different format. Real requests for help aren't public, so we wrote those reports ourselves based on real places in Buncombe County.

Accomplishments that we're proud of

We got the full flow working, from someone texting for help to a volunteer showing up and closing out the task. We're also proud of the risk map. We tested it against the landslides USGS mapped after Helene, and the 20% of the county we rated highest risk had about twice as many landslides as the rest of the map would predict.

What we learned

We learned how to build a flood and landslide risk map from public data, using things like height above the nearest stream, slope, and rainfall. We also got a lot more comfortable with PostGIS, especially routing on a road network and removing closed roads from it. For linking reports, we ended up using a SQL query to find nearby incidents from the past 48 hours, then letting the LLM decide whether the report was about one of them. Running the replay on a simulated clock also mattered, so the system only knew what had been reported up to that point.

What's next for FieldLine

Extend FieldLine as a platform where people can directly create requests for aid and communicate during the crisis
Integrate more data from satellites, drones, and street view cameras