A Hybrid Edge-AI Solution Combining Large Language Models and Computer Vision for Enhancing AgTech Access Among Indian Smallholder Farmers

Abstract

Agricultural advisory systems find it particularly challenging to reach smallholder farmers in rural and low-resource areas of India, who have a variety of linguistic needs, low literacy, and limited internet access. In this paper, we present AgriLLM X, a novel Edge-AI framework that provides intelligent, localized, and explicable agricultural support in real time by combining large language models (LLMs), computer vision, multimodal fusion, and genomic data analysis. AgriLLM-X combines contemporary AI methods such as low-rank adaptation (LoRA) for optimizing LLMs on agricultural corpora and retrieval augmented generation (RAG) for context-aware knowledge retrieval. There are three main modules: the EVSF (Edge Vision-Sensor Fusion) module, for real-time diagnosis using images and sensor readings, MLAS (Multilingual Local Advisory System) module for localized voice dialog in regional languages, and the GETA (Genomic Trait Analyzer) module for recommendations based on genes. With field based multimodal data collection and optimization techniques consisting of model quantization and pruning employed for energy-efficient edge deployment, development was determined by a systematic empirical methodology. The F1-score on genomic trait prediction (89%), voice recognition (92.3%), and plant disease identification (96.2%) indicate good performance of the system on all evaluated aspects. Significant real-world effects were also observed, as 15–22% increase in yield, a 36% increase in access for female farmers, and a high user satisfaction rate of 91%. AgriLLM-X is a scalable, modular, and explicable solution designed for AI-enabled, inclusive, and climate-resilient agriculture. Its deployment model provides a straightforward, replicable strategy for smart farming in places with poor internet connectivity.