🔥 FireSight — Every spark in sight.
Early wildfire detection, worldwide, free — often before the media.
🌍 Submitted to DreamHacks 2026 🌱
🐤 Inspiration
The canary once sang in the coal mine — a tiny signal that saved lives because it came early.
Wildfires today move faster than the information about them. Again and again, the same pattern repeats:
| 😟 The painful reality | 💥 What it costs |
|---|---|
| 🕐 Official alerts often arrive hours after ignition | Families lose the "golden hour" to evacuate |
| 📰 Most people learn about fires from the news | The story breaks after the fire is out of control |
| 🛰️ Satellites see everything — but the data is raw, scattered, expert-only | A life-saving signal sits unread in a NetCDF file |
| 🗣️ Eyewitnesses post within minutes on social media | Their warnings drown in noise, unverified and unmapped |
| 💸 Good fire dashboards are paywalled or regional | Most of the planet has no free, live picture at all |
We asked a simple question: what if the planet had a canary? A system that never sleeps, watches every satellite pass, listens to every verified witness, and taps you on the shoulder the moment a spark appears — in your language, for free.
That question became FireSight.
🔥 What it does
FireSight is a free, independent, near-real-time world map of wildfire ignitions. One glance tells you what is burning, where, how sure we are, and who is fighting it.
| 🧩 Capability | ✅ What you get |
|---|---|
| 🌐 Live fire globe & map | Every active ignition on Earth on an interactive 3D globe (MapLibre GL), with 6 h / 24 h / 72 h time windows, a smoke layer, city search and "My location" |
| 🛰️ Multi-satellite fusion | NASA FIRMS (VIIRS 375 m) + NOAA GOES-18/19 (read directly from public AWS buckets, ~10-minute refresh — about 30 minutes earlier than FIRMS) + EUMETSAT Meteosat MTG, merged and de-duplicated |
| 🤖 Double AI verification | Citizen reports from Bluesky, press (GDELT) and Telegram are geoparsed, then judged twice by independent AI models — "is this a fire right now?" and "which place is the fire?" — before they ever touch the map |
| 🚦 Honest confidence levels | Every fire climbs possible → probable → corroborated as evidence stacks up. No crying wolf |
| ✈️ Water-bomber tracking | Firefighting aircraft identified live from ADS-B (registration + ICAO type) — watch the response unfold |
| 🔔 Free area alerts | "Alert me on this area" — one tap, web-push notifications, no account, checked every 5 minutes |
| 📖 Fire memory | Every significant fire gets a permanent, citable archive page — a memory the internet keeps |
| 📊 Observatory | Citable wildfire statistics per country and per month, for journalists, researchers and policymakers |
| ⏱️ Measured earliness | We compare our first detection against first press coverage — and publish our lead time |
| 🌍 For everyone | Fully translated in French, English, Spanish and Portuguese, installable as a PWA, embeddable as a free widget |
| 🤝 Agent-native | A public MCP server lets Claude, ChatGPT, Cursor or any AI agent query live fire data as first-class tools |
⚠️ FireSight is an information service, not an official alert channel. In an emergency, call 112 (Europe), 911 (North America), 193 (Brazil) or your local emergency number.
📸 See it in action
*🌍 The live 3D globe — every active ignition on Earth. 15 firefighting aircraft airborne. 116 active fires in Russia at capture time.*
*🕐 The 24-hour window over the Americas — 254 active fires in Brazil, tracked in real time.*
*🔎 Zoomed into South America — clustered fire events, live per-country counters, one-tap area alerts.*
🛠️ How we built it
FireSight is a detection pipeline, not a dashboard. Raw space hardware telemetry goes in one end; a trusted, human-readable fire alert comes out the other.
┌──────────────────────────────────────────────────────────────────────────────────────────────────┐
│ 1. INGESTION & INPUT TIER │
├──────────────────────────────────┬─────────────────────────────────┬─────────────────────────────┤
│ 🛰️ EYES IN THE SKY │ 🗣️ EYES ON THE GROUND │ 📡 FLIGHT TRACKING │
│ • NASA FIRMS (VIIRS 375m) │ • Bluesky Social Feed │ • ADS-B Firefighting │
│ • NOAA GOES-18/19 (AWS S3) │ • GDELT Global Press │ Aircraft Telemetry │
│ • EUMETSAT MTG (Meteosat) │ • Telegram Chats & User Reports│ │
└────────────────┬─────────────────┴────────────────┬────────────────┴──────────────┬──────────────┘
│ │ │
▼ ▼ │
┌───────────────────────────────────────────────────────────────────────────────────┼──────────────┐
│ 2. CORE PROCESSING PIPELINE │ │
├───────────────────────────────────────────────────────────────────────────────────┤ │
│ 🧠 FireSight Core Engine (Next.js 16 + TypeScript Architecture) │ │
│ │ │
│ ┌──────────────────────────────┐ ┌──────────────────────────────┐ │ │
│ │ 📥 Satellite Ingestion │ │ 🌍 Gazetteer Geoparsing │ │ │
│ │ (WASM HDF5/NetCDF Parsing) │ │ (Location Resolution) │ │ │
│ └──────────────┬───────────────┘ └──────────────┬───────────────┘ │ │
│ │ │ │ │
│ ▼ ▼ │ │
│ ┌──────────────────────────────┐ ┌──────────────────────────────┐ │ │
│ │ 📍 Spatial Event Clustering │ │ 🤖 Double AI Verification │ │ │
│ │ (Grid + Connected Comp.) │ │ (2 Independent LLM Judges) │ │ │
│ └──────────────┬───────────────┘ └──────────────┬───────────────┘ │ │
│ │ │ │ │
│ └─────────────────┬──────────────────┘ │ │
│ │ │ │
│ ▼ │ │
│ ┌───────────────────────────────────────────────────────────────────┐ │ │
│ │ 🔥 Confidence Grading Engine (Possible ➔ Probable ➔ Corroborated)│ │ │
│ └────────────────────────────────┬──────────────────────────────────┘ │ │
│ │ │ │
│ ▼ │ │
│ ┌───────────────────────────────────────────────────────────────────┐ │ │
│ │ 🗄️ Fire Memory Observatory & Precocity Metrics (Earliness Index) │ │ │
│ └────────────────────────────────┬──────────────────────────────────┘ │ │
└───────────────────────────────────┼───────────────────────────────────────────────┼──────────────┘
│ │
▼ │
┌───────────────────────────────────────────────────────────────────────────────────┼──────────────┐
│ 3. DATA STORAGE LAYER │ │
├──────────────────────────────────────────────────┬────────────────────────────────┤ │
│ 🗄️ Supabase Postgres DB │ 📦 Vercel Blob Cache │ │
│ • Historical Fire Archives │ • Shared Geometry Cache │ │
│ • Contextual Signal Stores │ • Pipeline Interim Arrays │ │
│ • System Analytics & Trends │ │ │
└────────────────────────┬─────────────────────────┴────────────────┬───────────────┘ │
│ │ │
└──────────────────┬───────────────────────┴───────────────┐ │
│ │ │
▼ ▼ ▼
┌──────────────────────────────────────────────────────────────────────────────────────────────────┐
│ 4. PRESENTATION & DELIVERY TIER │
├───────────────┬───────────────┬────────────────┬─────────────────┬────────────────┬──────────────┤
│ 🗺️ LIVE GLOBE │ 🔌 REST API │ 🤖 MCP SERVER │ 🔔 ALERTS │ 🧩 EMBEDS │ 📦 OPEN DATA │
│ MapLibre GL │ JSON, CSV, │ Claude, Cursor │ Real-time Web │ Drop-in Canvas │ CC BY 4.0 │
│ & Mobile PWA │ & RSS Feeds │ & ChatGPT │ Push Subsystem │ Widgets │ Public Pool │
└───────────────┴───────────────┴────────────────┴─────────────────┴────────────────┴──────────────┘
🧬 The journey of one spark
🛰️ Satellites ⚙️ Ingestion 🧠 Clustering 🤖 AI Judges 🗄️ Archive 👤 You
│ │ │ │ │ │
1. │── Thermal ──────►│ │ │ │ │
│ Detections │ │ │ │ │
│ │ 2. Clean │ │ │ │
│ │─── GeoJSON ─────►│ │ │ │
│ │ │ │ │ │
│ │ │ 3. Run ~4km │ │ │
│ │ │──┐ Grid Cluster │ │ │
│ │ │ │ │ │ │
│ │ │◄─┘ │ │ │
│ │ │ │ │ │
│ │ │ │ [Meanwhile: │ │
│ │ │ │ Ground Feeds] │ │
│ │ │ │ │ │
│ │ │ │ 4. Run Judge 1 │ │
│ │ │ │──┐ (Is it a │ │
│ │ │ │ │ live fire?) │ │
│ │ │ │◄─┘ │ │
│ │ │ │ │ │
│ │ │ │ 5. Run Judge 2 │ │
│ │ │ │──┐ (Where is │ │
│ │ │ │ │ the fire?) │ │
│ │ │ │◄─┘ │ │
│ │ │ │ │ │
│ │ │ 6. Attach │ │ │
│ │ │◄── Verified ─────│ │ │
│ │ │ Witnesses │ │ │
│ │ │ │ │ │
│ │ │ 7. Archive Event │ │ │
│ │ │────────────────────────────────────►│ │
│ │ │ │ │
│ │ │ │ 8. Broadcast │
│ │ │ │─── Map/Alert/───►│
│ │ │ │ API Updates │
🧰 The stack
| Layer | Technology | Why it's cool 😎 |
|---|---|---|
| 🖥️ Framework | Next.js 16 (App Router) · React 19 · TypeScript 5 | Server components + route handlers run the entire public API |
| 🗺️ Maps | MapLibre GL 5 | GPU-accelerated globe projection with custom flame sprites |
| 🛰️ Satellite parsing | h5wasm · fflate | Raw NetCDF4/HDF5 GOES fire products parsed in pure TypeScript/WASM — no Python, no GDAL, no server-side binaries |
| 🤖 AI | OpenAI — two independent judging passes + custom gazetteer geoparsing | Consensus before any human report reaches the map; kills classic traps like "The Globe and Mail" → Globe, Arizona |
| 🗄️ Data | Supabase (Postgres) · Vercel Blob | Permanent fire memory + shared edge caches |
| 🤝 Agent protocol | Model Context Protocol (mcp-handler) |
LLMs query live wildfire data as native tools |
| 🔔 Notifications | Web Push (VAPID) | Per-area subscriptions dispatched by a 5-minute cron |
| ⚙️ Ops | GitHub Actions (CI + cron) · Vercel | Zero-cost scheduler; CI builds and validates everything with zero secrets |
| ✅ Quality | Zod 4 · custom SEO guard · QA harness | Every external payload schema-checked; structural SEO contracts fail the build |
📈 By the numbers
| 📄 TS/TSX files | 🧮 Lines of code | 🛰️ Satellite sources | 🗣️ Witness channels | 🌐 Languages | ⏱️ Refresh | 💰 Price |
|---|---|---|---|---|---|---|
| 124 | ~23 900 | 3 agencies · 5+ instruments | 3 | 4 | ~10 min | $0, forever |
🧗 Challenges we ran into
| 🧩 Challenge | 🛠️ How we solved it |
|---|---|
| FIRMS publishes GOES data ~40 min late | We bypassed it: read the raw ABI L2 fire products straight from NOAA's public AWS buckets — available ~10 min after scan. That meant parsing binary HDF5/NetCDF in TypeScript via WebAssembly (h5wasm), including a nasty silent-NaN bug where scalar HDF5 attributes arrive as TypedArray(1) and poison every geolocation calculation |
| Social media is a firehose of false alarms | One AI pass wasn't enough. "The Globe and Mail" is a newspaper, not Globe, Arizona. So every candidate post is judged by two independent AI models — one for relevance ("is this a fire happening now?"), one for place disambiguation — with verdicts cached for 7 days to control cost |
| Turning dots into fires | Individual hotspots are noisy. We built a stateless spatial-grid + connected-components (8-neighbour) clustering engine that groups detections into events in one pass — the first detection becomes the ignition-time proxy |
| Free tier, 5-minute cadence | Serverless crons only run daily — so GitHub Actions became our free 5-minute trigger, with timeouts aligned to function limits so slow scans don't false-alarm |
| Truthful UI under uncertainty | Fires are scary; false alarms are worse. The whole UI speaks in confidence levels, not absolutes, with a permanent emergency disclaimer and measured self-evaluation (we publish our lead time over the press) |
🏆 Accomplishments that we're proud of
- ⏱️ We beat the press — and prove it. The precocity module measures how much earlier FireSight surfaces each fire versus first media coverage, and publishes the number.
- 🛰️ GOES data ~30 minutes earlier than the standard feed by reading space-agency hardware products directly — in the browser-era stack, with zero native dependencies.
- 🤖 A double-AI jury for citizen reports that actually works in production, with cached verdicts and QA re-judgement to catch regressions.
- 🤝 One of the first wildfire datasets with an MCP server — any AI agent on Earth can ask FireSight "what's burning near me?" in one tool call.
- 💪 Zero-secret resilience: every page degrades gracefully, and CI proves it on every push with a blocking SEO guard, type gate and production build.
- 🌍 Free, open, multilingual — 4 languages, CC BY 4.0 data, no accounts, no paywalls. The communities most at risk pay nothing.
📚 What we learned
- 🧊 Deep satellite engineering: VIIRS vs geostationary trade-offs, FRP (Fire Radiative Power), detection confidence codes, and the quirks of HDF5 attributes in WASM.
- 🤖 AI in production is a systems problem: prompt design matters less than caching, batching, budgets, and knowing when not to call the model (keyword-only fallback when no key is present).
- 🗺️ Geospatial algorithms: grid clustering, connected components, gazetteer matching, perimeter and risk layers (NIFC, DFCI).
- 🔔 Web Push end-to-end: VAPID, service workers, subscription storage, and cron-based dispatch.
- 🤝 Model Context Protocol: designing tools that LLM agents can use safely and usefully.
- 🧭 Product honesty: how to communicate uncertainty (
possible→probable→corroborated) without either panic or false comfort.
🚀 What's next for FireSight
- [ ] 📡 More eyes in the sky — Sentinel-3 SLSTR and additional geostationary feeds
- [ ] 🌫️ Smoke-drift forecasting — where will the smoke be in 6 hours?
- [ ] 🏘️ Community verification network — partnering with local environmental organizations and fire services (as the hackathon theme encourages!)
- [ ] 📱 Native mobile app — the PWA is live; native push is next
- [ ] 🧾 DOI for the open dataset — making FireSight citable in academic research
🏅 Why FireSight — judging criteria at a glance
| 🥇 Criterion | 🔥 How FireSight answers |
|---|---|
| Originality | Not another dashboard — a full detection pipeline: direct-from-AWS satellite parsing, dual-AI-verified citizen signals, live firefighting aircraft, permanent fire memory, and a published lead-time metric |
| Adherence to Track | Earth Forward, fully 🌱: earlier detection = faster response = saved forests, saved biodiversity, avoided CO₂, saved lives — delivered free to the communities that need it most |
| Completion | Everything works: live globe, alerts, archive, observatory, REST/CSV/RSS API, MCP server, embeddable widget, PWA, 4 languages — enforced by CI on every push |
| Learning | HDF5-in-WASM, MCP, ADS-B decoding, geospatial clustering, web-push — all new ground, all in one build |
| Design | A real design system (handoff/) with brand tokens and a tone charter — calm, precise, never alarmist; clean mobile-first UI |
| Technology | 3 satellite constellations + 3 witness channels + 2 AI judges + ADS-B + push + MCP — many hard parts, one elegant pipeline 🤯 |
🔗 Links
- 💻 Source code: github.com/theyapguard/FireSight
- 📦 JS client:
packages/firesight-fires - 🤖 MCP manifest:
server.json
### 🌱 Earth Forward — because the planet can't wait for the news. 🔥 **Built with ❤️ for DreamHacks 2026**
Built With
- eumetsat-meteosat-mtg
- fflate
- github-actions
- h5wasm
- maplibre-gl-5
- model-context-protocol-(mcp)
- nasa-firms
- next.js-16
- noaa-goes
- openai-api
- openstreetmap
- react-19
- supabase
- tailwind-css
- typescript-5
- vercel-blob
- web-push-api-(vapid)
- zod-4
Log in or sign up for Devpost to join the conversation.