PROJECT INTEL
Iran Wildfire Detection
Production-grade real-time wildfire detection for Iran — a fault-tolerant multi-stage agent pipeline.
- ACTIVE SINCE:
- 2025 — present
- STATUS:
- ACTIVE
- FIREPOWER
- 8/10
- ARMOR
- 9/10
- SPEED
- 9/10
- SPECIAL
- 7/10
A real-time agent that watches Iran's forests so nobody has to refresh NASA dashboards all day. A custom fault-tolerant DAG executor handles dependency resolution, parallel task execution, per-node timeouts, exponential backoff and circuit breakers — degrading gracefully when individual data sources fail. It fuses NASA FIRMS hotspots from three VIIRS satellite products with ESA WorldCover land data, Open-Meteo and OpenWeather — coordinating five-plus external APIs with very different latencies, rate limits and reliability.
Severity is scored by a hybrid LLM + rule-based system — structured prompting, JSON schema validation, and automatic fallback to rules when the LLM fails — combining fire radiative power, fuel availability, weather and air quality into evidence-based assessments. A single async thread juggles 100+ concurrent API calls; DAG parallelisation cut end-to-end latency by roughly 60%, and ACID-backed SQLite state keeps runs idempotent, deduplicated and fully audited.
BATTLE RECORD
- Fault-tolerant DAG executor: per-node timeouts, backoff, circuit breakers
- Multi-source fusion: 3 VIIRS satellite products + land cover + weather APIs
- Hybrid LLM + rule-based severity scoring with automatic fallback
- 100+ concurrent API calls on a single async thread; ~60% latency cut
- ACID-backed state: idempotent runs, deduplication, full audit log
TECH
- Python
- asyncio
- DAG executor
- NASA FIRMS
- ESA WorldCover
- Hybrid LLM + rules
- SQLite