AGRO MIRAI reads a field's soil, weather, and satellite signals to ground every recommendation in real data, not a guess — and speaks it back in the farmer's own language.
Every recommendation is generated from the farmer's own field data — soil samples, local weather, and satellite NDVI — not a generic lookup table.
Ranks all 22 supported crop types against FAO EcoCrop's own temperature, soil-pH, and soil-texture ranges — plus the field's real 12-month rainfall and a sowing-month calendar. A regional-suitability check flags any pick that's agronomically unusual for the farmer's district, instead of silently trusting the model.
A daily FAO-56 root-zone water balance — Hargreaves-Samani reference evapotranspiration (chosen over Penman-Monteith, since a real farm doesn't reliably report solar radiation or wind), crop coefficient by growth stage, root depth, and soil water-holding capacity. The same numbers set both the recommended depth in mm and the urgency — never two models that can disagree.
A MobileNetV2 CNN trained on the PlantVillage dataset reads an uploaded leaf photo, when one's available. When it isn't, a weighted rule-based score over humidity, rainfall, temperature, and NDVI trend fills in — and if the CNN call itself fails, that same rule-based score catches it. Never a dead end.
Every advisory translates and speaks back in English, Kannada, Telugu, and Hindi (IndicTrans2 + AI4Bharat ASR/TTS, Piper for English and Hindi) — and login is just name, phone, and an OTP, so there's no password for a farmer to forget.
The same pipeline runs behind every advisory type — only the model at the processing step changes.
Every figure on this page comes from a real, committed evaluation report — nothing hand-faked.
A React Native / Expo mobile app puts all four advisories, voice playback, and photo-based disease scans on one screen a farmer already knows how to use — no password, just name, phone, and an OTP.
Direct install — no Play Store yet. Android will ask to allow "install from unknown sources". Release notes & checksum.
Where a figure comes from a trained model's held-out evaluation, we say so. Where it comes from a rule or a real dataset instead, we say that too — crop and irrigation run on transparent methods, not a black-box accuracy score.