AI Hazard Finder: Real-Time Extraction of Time, Location, and Type for Global Hazards

Timely, high-fidelity situational awareness requires extracting what/when/where from high-volume news streams. We present AI Hazard Finder (HF), a production two-stage pipeline that couples lightweight rules/heuristics with confidence-triggered large language model (LLM) reasoning to jointly infer hazard type, best-estimate time, and location with calibrated scores. Stage 1 applies a hazard-type classifier, rule-augmented temporal tagger, and geoparser to process ≈ 20,000 articles per week; Stage 2 escalates only low-confidence items to a LLaMA-3 8B extractor for joint correction and completion. On a manually annotated set (n=249), the system achieves weighted F1 = 0.617 (hazard type), 0.594 exact match for time (0.659 within ±6h), and 0.442 exact match for location. Deployed within PDC's DisasterAWARE platform, the pipeline currently triages ≈ 2,860/day and reduces analyst review time per geopolitical hazard by ~15 minutes. The annotated test set and evaluation scripts are available upon request via email to the authors; the production codebase remains proprietary.