Investigated withNVIDIA NemotronNebiusTavily
Snap a leaf and say what’s wrong in Hindi or English. We cross-examine it against your field’s past week of weather, Sentinel-2 satellite imagery and a rule-backed evidence score, have NVIDIA Nemotron check it against official agricultural guidance, then tell you what’s likely, what isn’t, and what to do next.
Under the hood
What happens in the minute after you tap send.
Not one model taking a guess. A pipeline that gathers evidence from your photo, your voice, the sky and space, scores it with published agronomy, has a second model cross-examine it, and shows its working.
Crop ID with a confidence gate
A multimodal model names the crop from your photo across 12 supported crops. Below the confidence bar, it asks you instead of guessing.
12 crops · ask-if-unsure
Symptom read
Spots, margins, halos, colour shifts and where on the plant they sit, recorded as observations and kept apart from any hypothesis.
observed ≠ suspected
Hindi and English speech
Your voice note is transcribed with the language detected automatically. Hindi stays in Devanagari; nothing gets translated away.
auto-detect · ~2 s
Seven days of field weather
Rain, humidity and temperature at your field’s coordinates from Open-Meteo, matched against the conditions each disease favours.
Open-Meteo · daily
Sentinel-2 canopy check
Cloud-masked Sentinel-2 L2A scenes over a 100 m buffer around your field, via Google Earth Engine: NDVI and NDMI trends across 45 days.
NDVI · NDMI · 45 days
An evidence ledger
Each cause gets a 0–100 fit: the model’s read plus bounded, rule-based weather and satellite points, each with its reason. Weak fits are dropped.
rules, not a black box
Triage, not a verdict
Urgency is rated, treatments are flagged for expert confirmation, and missing data is never quietly treated as normal.
expert-check flags
Five languages, zero storage
Results in English, हिन्दी, ਪੰਜਾਬੀ, Español and Français. Photos and voice notes are analysed in memory and never stored.
in-memory · 5 languages
Stage 2 · Investigation
Once the evidence is scored, a second model cross-examines it against official guidance. It can question the result, never rewrite the scores.
Official guidance, approved publishers only
Tavily searches research and extension publishers such as ICAR, state agricultural universities and FAO for the leading causes. Every page is re-checked for your crop and the disease, and dose instructions are stripped out.
allowlisted publishers · cached 6 h
A second opinion that shows its working
NVIDIA Nemotron 3 Ultra on Nebius Token Factory reads the evidence ledger and the sources, names what fits and what doesn’t, and asks the one question that would settle it; Nemotron 3 Super updates the review once you answer. Guardrails drop anything uncited or unsafe.
text only · cannot change a score
The RootCause method
A triage, not a verdict.
01
Observed
What your photos and words actually show, kept apart from any guess.
02
Suspected
Possible causes with their confidence, and the evidence that argues against them.
03
Missing
What would make the picture clearer, so you know exactly what to check.
04
Next steps
Safe actions ranked by urgency, flagged when an expert should confirm.
Two minutes, three steps
Photograph
Up to three close-ups of the affected leaves, stems or fruit.
Describe
Just say what you see, in English or Hindi. No typing needed.
Triage
See how urgent it looks, what it might be, and what to do next.
Something wrong in the field?
Send a report now. It works on any phone and never claims more certainty than the evidence allows.
Start a crop report