A production-grade fire intelligence platform covering the entire Brazilian Amazon, built solo as my IEEE journal internship and major project. An 8-step pipeline fuses INPE fire detections, ERA5 weather reanalysis and Google Earth Engine satellite features into a 200K-row dataset, on which 7 models — Random Forest, XGBoost, LightGBM, CatBoost, TabNet, ANN and a stacked ensemble — are trained on GPU, reaching 0.987 ROC-AUC. Every prediction is explained live with SHAP, fire spread is simulated with the full Rothermel physics model (the same family used by FARSITE), and a GIS module scores ecological risk to protected areas, indigenous territories and settlements. It all runs behind a FastAPI backend and a real-time Next.js + Leaflet dashboard with NASA FIRMS active fires, live weather heatmaps and an admin panel — plus a 5-agent DeepSeek LLM system that autonomously coordinates fire-crew interceptions, saving up to 30% of burn area in IEEE evaluation scenarios for under a cent per run.