VerdantAI

Verdant AI is a crop-health advisor for smallholder farmers in Ghana, built as a mobile-first web app. A farmer photographs a maize or cassava leaf, or a cocoa pod, and gets back the measured leaf with candidate lesions marked, the affected area as a percentage, a plain explanation of the likely cause and alternatives, what to do this week, when to call an extension officer, and a saved record. Advice comes in English, Spanish and a draft Twi, and can be read aloud. Ghana has about one extension officer for every 1,200 farmers (GNA), so a disease or fall armyworm outbreak can spread before anyone qualified sees it. I'm building Verdant Foods, a regenerative agriculture venture in Ghana, and this is the gap I see from the farm side. The app is built for prepaid data and weak signal. It opens offline, keeps photos on the phone and sends them when the connection returns (with an idempotency key, so a retry never causes a second paid scan), shrinks photos to 1280 px before upload, and sends spoken advice as 32 kbps MP3 (a Twi message drops from 1.6 MB to 203 KB). Farmers sign in with a phone number and PIN, the same pattern as mobile money, or continue as a guest. Beyond the scan there is a 7-day rain outlook for the farmer's region, an opt-in nearby alert that only shows a pest once at least 3 farms report it within 14 days, and follow-up questions about a result answered in the farmer's language. Measured on a held-out test set of 2,760 images, the EfficientNet-B0 classifier reaches 96.1% accuracy and 0.955 macro-F1 across 16 classes, and 98.1% accuracy on the 92.9% of images where it doesn't abstain. A real Muse Spark scan takes about 14 s and costs about $0.01. Total Muse Spark spend during the build was $0.051. 179 tests pass. The app gives a preliminary assessment, never a diagnosis. It never names a chemical product or dose, and it sends uncertain or severe cases to an extension officer. No field trials or farmer interviews have been done yet; the README lays out the pilot that would come next.