Firetower: off-grid triage for when the towers go down

When a storm takes out cell towers and internet, people still need to report emergencies, and someone still has to decide what to act on first. After Hurricane Helene, one North Carolina county received about 11,000 requests from people trying to find loved ones and had to sort them by hand.

Firetower lets people send short texts over low-cost Meshtastic LoRa radios (no cell service or internet needed). Each message reaches a gateway computer (a laptop for this prototype), where an open-weight model (OpenAI’s gpt-oss-20b, served by DigitalOcean) ranks it as EMERGENCY, PRIORITY, WELFARE or ROUTINE and pulls out where it happened, what happened and how many people are hurt. The sender gets a ticket number back over the radio, every reply is confirmed delivered, and a coordinator board shows every report sorted by urgency as it comes in.

When a report is missing a location or injury count, Firetower asks one short follow-up question over the radio and adds the answer to the same ticket. If the sender’s radio has no GPS, it estimates where they are from the place they named, the radio’s last known position, or which radio relayed the message, and marks the pin as an estimate. A keyword safety net catches urgent messages the model ranks too low (words like trapped, bleeding or insulin), and messages from priority nodes (shelters, households with a registered medical need) are never ranked below Priority. A ticket closes only when the sender says the situation is over, so “nobody hurt” never closes an active fire.

When the coordinator dispatches fire, police, EMS or rescue, Firetower radios the reporter which teams are coming, and the map shows nearby fire stations, hospitals and police stations with a simulated drive from the nearest one. Firetower never dials 911 or contacts a department itself; a person makes that call.

With the internet unplugged, Firetower keeps taking reports: the keyword net ranks urgency, tickets are stored on the gateway, and the board keeps working. The model’s extraction and follow-up wording return when the link does.

In a real deployment, the laptop would be replaced by a rugged, low-power computer such as an NVIDIA Jetson Orin, running on solar and battery power. It would sit at a fixed site like a fire station or shelter, with a high-gain antenna on a mast for better range. Because gpt-oss-20b is open-weight, a box like that could also run the model on site, so full triage works with no internet at all.

Built with: Meshtastic, Python, FastAPI, SQLite, DigitalOcean (gpt-oss-20b), MapLibre, OpenStreetMap (Nominatim, Overpass, OSRM).