Delphi
Predicting maize pest outbreaks before the damage, with explanations, not black boxes.
★ Winner · AI & Automations, HackITBA 2026
Overview
A predictive-intelligence platform that combines population-monitoring data, climate variables, and geospatial signals (satellite imagery, terrain) to estimate pest outbreak risk by locality and make that information actionable. Delphi anticipates the maize leafhopper (Dalbulus maidis) before it damages crops, turning early warnings into planting decisions. It's built on real data from Argentina's INTA National Trap Monitoring Network (330+ traps, 38 biweekly reports across two seasons) paired with a pre-season risk score and 14-day tactical monitoring that reads recent captures, weather, and propagation across neighboring regions. Every prediction is explainable through SHAP feature contributions, visualized in an interactive map with locality-level drill-down and AI-generated advisories.
Highlights
- Won the AI & Automations category at HackITBA 2026.
- Two-horizon modeling: pre-season XGBoost risk scoring plus biweekly tactical monitoring with neighborhood propagation signals.
- SHAP-based explainability surfaces locality-specific risk drivers instead of a single opaque number.
- Full product: FastAPI backend on Railway, Next.js 16 and React 19 frontend on Vercel, Supabase auth.