HackBarna 2026 · 19–20 Sep · Norrsken House Barcelona
HackFire
A voice agent that calls residents before a wildfire reaches them, tells them how to get out, and shows the emergency coordinator who needs rescue.
Norrsken × Deepfire “AI for Wildfire” challenge, track 4: values at risk.
6 h 08 min
early warning
calling ahead of the fire
- 6 h 08 minLa Atalaya flagged before the first satellite hotspot came within 3 km, from satellite data only.
- 7,068Deepfire satellite hotspots from 22–24 July 2026, replayed on a time slider.
- 6 toolsThe HTTP contract the voice agents use: fire status, routes, triage, rescue queue, crew plan.
- 60–100 sReal outbound calls with the resident agent, logged by SLNG on 20 September.
The problem
On 23 July 2026 the Burgohondo wildfire in Ávila ran 3 km in 40 minutes and reached the La Atalaya housing estate in El Tiemblo. Between 1,200 and 1,300 people were evacuated and 5 homes were destroyed.
Public alerting is one-way. ES-Alert broadcasts the same message to a whole area and doesn’t listen: the coordinator can’t tell who has left, who never got the message and who can’t move. Rescues surface late, through saturated 112 lines.
“Every resident who evacuates on their own is a rescue the firefighters don’t have to make.”
How it works
one household, end to end
- 1The fire and where it’s headingDeepfire satellite hotspots on a map with a time slider, plus an hour-by-hour predicted spread. Live mode shows fires burning now.
- 2Who is in its pathThe spread is intersected with towns, housing estates, care homes and roads from OpenStreetMap. Each zone gets a time to impact.
- 3The agent calls residentsA route out that avoids the predicted fire, then three questions: can you leave on your own, how many are you, what can you see.
- 4Triage for the coordinatorEach resident becomes evacuating, no answer or needs rescue. Rescues are ranked by time to impact and texted to the fire crew.
What’s real, what’s demo
Real
- Deepfire’s satellite hotspots of 23 July 2026, and in live mode the fires burning now in Spain.
- Places and roads from OpenStreetMap; routes from openrouteservice.
- The voice agents: real AI agents on SLNG, on NVIDIA Nemotron Super 3.
Fictional, for the demo
- The residents, their homes and what they answer.
- The calls in the replay, the evacuations and the crew alerts.
- Nobody near the real fire was called — the dashboard says so on screen.
My role
We planned the idea and built it with Claude Code, using Matt Pocock’s skills: a grill-me session to pin the plan down, the work split into vertical-slice issues built test-first so each one worked end to end, and /design for the design.
Built with
- Python · FastAPI
- React · TypeScript
- MapLibre GL
- NVIDIA Nemotron Super 3
- SLNG voice agents
- Deepfire API
- openrouteservice
- OpenStreetMap
- Vonage SMS
- Railway
Shall we go on with the rest?