Pyroscope

Inspiration

During wildfire season in Canada, wildfire analysts receive hundreds of satellite thermal alerts daily. Most are smoke or false positives. But buried in that noise are fires that can threaten lives within hours. Manual triage is slow, inconsistent, and exhausting. We wanted to build a tool that cuts through the noise instantly and tells analysts exactly where to look first.

What it does

Pyroscope is a real-time wildfire triage dashboard. It pulls live satellite hotspot data from NASA FIRMS, displays all active fire detections on an interactive map, and lets analysts request an AI-powered risk assessment for any hotspot on demand. Each analysis combines the satellite image of the location, live weather conditions (wind speed and direction, humidity, temperature), and fire radiative power (FRP) to produce a risk rating (high, medium, low), along with terrain analysis, spread direction, and recommended actions.

How we built it

• Next.js 16 for the full-stack framework, where API routes handle data fetching server-side, React for UI
• NASA FIRMS API for real-time VIIRS satellite hotspot data across any region
• Open-Meteo for live per-coordinate weather data
• Mapbox Static Images API to pull a satellite image of each hotspot location
• Backboard.io as the AI backbone, it receives the satellite image alongside structured weather and fire data, and returns a structured JSON risk profile
• Leaflet for the interactive map with colour-coded risk markers

Challenges we ran into

The biggest challenge was prompt calibration. Our first prompt had the AI treating every low-intensity agricultural burn as an imminent catastrophe. It was recommending mass evacuations for a 2 MW field. We had to engineer the prompt to include scientific FRP context so the model could self-calibrate rather than default to worst-case assumptions. Also, finding active fire data was harder than expected. We cycled through Canada, Florida, and Australia before realizing it was peak dry season in sub-Saharan Africa, where 93% of global hotspots were concentrated yesterday.

Accomplishments that we're proud of

The pipeline is entirely real. We didn't use dummy data or hardcoded results. Every hotspot on the map is a live NASA satellite detection from the last 24 hours, every weather reading is pulled fresh per coordinate, and every risk assessment is a multimodal AI analysis of the actual satellite image of that location.

What we learned

FRP (Fire Radiative Power) is the key signal for calibrating fire severity. We also learned that AI risk tools need to be explicitly instructed to be conservative without guidance. Models lean toward dramatic responses, which in an emergency triage context is just as dangerous as missing a real threat.

What's next for Pyroscope

• Expand beyond Canada to a fully global, always-on view that follows fire season as it moves across hemispheres
• Add historical fire overlay to show spread over time
• Integrate wind vector visualization directly on the map
• Push high-risk detections to analysts via SMS or email alerts
• Explore finer-grained fuel moisture and drought index data to improve spread predictions

Built With

• backboard.io-sdk
• css
• leaflet.js
• mapbox
• nasa-firms-api
• nextjs
• node.js
• open-meteo
• react
• tailwindcss
• typescript