AgroGuard AI: A Zero-Shot Multimodal Crop Disease Diagnostic Assistant for Smallholder Farming

Smallholder farmers in India frequently incur avoidable
agricultural input costs due to non-diagnostic decisionmaking — replicating a neighboring farmer’s treatment schedule or acting on retailer-driven upselling —
rather than an actual, verified assessment of their crop’s
health, a behavioral pattern that often pushes vulnerable farmers toward avoidable debt. Existing plant disease detection systems, while accurate in controlled settings, remain research prototypes that assume technical
literacy and require dataset-specific training, offering
no low-friction mechanism to interrupt this cycle. This
work presents AgroGuard AI, a lightweight, browserbased diagnostic assistant that enables a farmer to independently verify crop health using only a photograph,
requiring no typing, technical training, or intermediary. The system encodes the captured image and a
role-conditioning prompt into a single multimodal request processed by Google’s Gemini API, which performs zero-shot visual reasoning — via patch-based
image segmentation and self-attention — to generate
a plain-language diagnosis, confidence estimate, and
treatment recommendation. Evaluation on a small
set of test images demonstrated preliminary diagnostic capability without any dataset-specific fine-tuning.
Critically, this work identifies and explicitly characterizes an important epistemic distinction largely overlooked in similar systems: the model’s stated confidence percentage is a generated linguistic estimate,
not a calibrated statistical probability as in a closedset classifier. While limitations around internet dependency, sample size, and visual explainability remain, the system demonstrates that accessible, meaningful agricultural decision support can be achieved
without heavy infrastructure or custom training data.