AgroSentinel an autonomous AI agent for agricultural water-stress monitoring, built for Costa Rica’s San Carlos farming region. It follows a perceive-decide-act loop: field data (NDVI, soil moisture, crop type, growth stage) and real NASA climate data feed into deterministic FAO-56 agronomic calculations (reference evapotranspiration, soil water balance, real crop coefficients cited from a Costa Rican engineering thesis), and only then does Claude reason over that already-validated science to decide a dosed irrigation valve’s exact opening percentage, with factor-by-factor reasoning and a stated confidence level. The full system includes a Supabase-backed multi-user database with row-level security, a Telegram bot for natural-language reporting and proactive alerts, and a multi-agent negotiation system where neighboring farms in the same watershed autonomously negotiate shared water through an AI mediator
Keeping the AI honest was harder than making it work. Early on, Claude was inventing plausible-sounding numbers (liters saved, days until crisis) with no formula behind them, we had to redesign the pipeline so all agronomic math is computed deterministically in code before Claude ever sees it, and Claude’s job became reasoning over validated data rather than generating numbers itself. That every scientific claim in the project is traceable to a real source
Sign up with email or Google. Field data can be entered manually via sliders, so no sensors or drone are required to try a full analysis, pick a crop and soil type, adjust NDVI/humidity/temperature, and hit “Analyze” to see Claude’s real decision.
