EmberWatch — Fire Detection & Intelligence System

EmberWatch is Prodigy's real-time wildfire intelligence layer. It continuously ingests fire data from10 government and satellite sources across 3 countries, normalizes it into a unified schema, and delivers property-specific threat assessments to homeowners — typically within15–30 minutes of fire detection.

The system runs autonomously 24/7 — no human operator. A scheduled job fetches, deduplicates, and caches fire data every 5 minutes. Users never hit external APIs directly; they read from a pre-built cache that delivers sub-200ms response times regardless of how many people are watching the same fire.

We pull from 10 live government feeds, organized into two parallel fetch batches to manage memory (each serverless function has a 128MB limit):

Every data source returns a different schema. The pipeline normalizes each fire into aunified record with consistent fields:

// Normalized fire record structure:
{
  fire_id:            "nifc_ABC123",        // Globally unique: source prefix + native ID
  region:             "US",                  // US | CA | AU
  source:             "NIFC",                // Which API provided this
  name:               "Oak Fire",            // Human-readable name
  latitude / longitude: 34.12, -119.45,     // WGS84 coordinates
  
  // Status normalization — every source maps to ONE of these:
  normalized_status:  "OUT_OF_CONTROL",      // OUT_OF_CONTROL | BEING_HELD | UNDER_CONTROL | UNKNOWN
  status_raw:         "OC",                  // Original value from source API
  
  // Classification
  is_prescribed:      false,                 // Prescribed/hazard reduction burns flagged separately
  is_controlled:      false,                 // 100% contained
  is_hotspot:         false,                 // Satellite thermal detection vs confirmed incident
  fire_type:          "WILDFIRE",            // WILDFIRE | PRESCRIBED | HOTSPOT
  containment_pct:    35,                    // 0-100 (null if unavailable)
  
  // Size & timing
  size_hectares:      5000,
  discovered_date:    "2026-03-15T08:00Z",
  
  // Metadata (varies by source)
  metadata: {
    agency: "USFS",                          // Jurisdictional agency
    cause:  "Lightning",                     // Fire cause if reported
    state:  "CA", county: "Ventura",         // US location detail
    province: null,                          // Canadian province
  }
}

With 10 sources, the same fire can appear multiple times (e.g., CWFIS and BCWS both reporting the same BC fire). We use fire_id as a composite key (source prefix + native ID) and merge into a HashMap — first-write wins.

Prescribed burns:        ≤ 7 days old
Controlled (100%):       ≤ 10 days old
NIFC incidents:          ≤ 30 days old    // US fires have slower reporting cycles
All other wildfires:     ≤ 21 days old

// Special case: NIFC fires with no discovery date are kept
// if they have size > 1ha OR containment < 100%

┌─────────────────────────────────────────────────────────────────┐
│  LAYER 1: FireCache Entity (Database)                           │
│  Updated every 5 min by refreshFireCache scheduled job          │
│  Contains ALL fires globally (typically 400-800 records)        │
│  Single row: cache_key = "global_fires"                         │
└──────────────────────────┬──────────────────────────────────────┘
                           │ DB read (once per 30s per isolate)
┌──────────────────────────▼──────────────────────────────────────┐
│  LAYER 2: In-Memory Cache (Deno Deploy Isolate)                 │
│  30-second TTL — shared across all concurrent requests          │
│  Request coalescing: 1000 simultaneous users → 1 DB read        │
│  Bounding-box pre-filter → O(relevant) not O(all)               │
└──────────────────────────┬──────────────────────────────────────┘
                           │ JSON response (~10KB per user)
┌──────────────────────────▼──────────────────────────────────────┐
│  LAYER 3: Client Session Cache (Browser RAM)                    │
│  Push-based reactive cache — updates on Layout prefetch          │
│  Instant restore on tab switch (0ms)                            │
└──────────────────────────┬──────────────────────────────────────┘
                           │ Fallback if offline
┌──────────────────────────▼──────────────────────────────────────┐
│  LAYER 4: localStorage Cache (5-min TTL)                        │
│  Enables offline mode — stale data better than no data          │
│  Auto-cleared when fresh data arrives                           │
└─────────────────────────────────────────────────────────────────┘

After the cache is written, the system runs a movement tracking phase. For every property with coordinates, it checks each nearby fire against previously recorded distances.

// Movement detection logic:
previousDistance = FireTracking.distance_km   // from 5 min ago
currentDistance  = haversine(property, fire)  // now
distanceChange   = current - previous

if (|distanceChange| < 0.5 km)  → "stable"
if (distanceChange < 0)         → "approaching"  // GETTING CLOSER
if (distanceChange > 0)         → "receding"     // MOVING AWAY

// If approaching:
approachRate = |distanceChange| / hoursSinceLastCheck   // km/h
estimatedArrival = currentDistance / approachRate        // hours

The EmberWatch map renders multiple overlay layers on Leaflet + Stadia Maps dark tiles:

• Fire Clusters — clustered markers at low zoom, individual fire icons at high zoom, color-coded by threat distance (red ≤5km, orange 5-15km, yellow 15km+, green = controlled)
• Monitoring Radius — adjustable circle (7/10/25 km) with traffic-light coloring (green = clear, yellow = elevated, red = critical)
• Fire Perimeters — real NIFC perimeter polygons when available, generated approximations from size for others
• Wind Overlay — animated directional arrows showing current wind speed and direction
• Fire Spread Prediction — projected spread zones based on wind + fire position
• Ember Attack Zones — modeled ember transport distance based on wind speed, fire intensity (FRP), and distance
• Evacuation Routes — AI-generated routes using OSRM road-following routing
• Historical Fire Footprints — past major fire boundaries for context
• Emergency Services — hospitals, fire stations, shelters from curated static database
• Road Closures — active closures near the property

Fire Movement Detected
        │
        ▼
  Alert Level Changed?  ─── No ──→  Skip (no spam)
        │
       Yes
        │
        ▼
  Is Prescribed Burn? ─── Yes ──→  Skip (never alert on Rx)
        │
       No
        │
        ▼
  Is Test Mode? ─── Yes ──→  Skip (demo safety)
        │
       No
        ├──→  In-App Notification (always)
        ├──→  Push Notification (native APNs / Firebase)
        ├──→  Email (Resend API — fallback or if enabled)
        └──→  SMS (Twilio — if phone verified & enabled)