Abstract: Despite significant advancements in computer vision, a gap remains between technological progress and real-world applications. This study aims to close this gap by providing farmers and technicians with an integrated framework that uses Deep Learning (DL) and Large Language Models (LLM) to enable early detection of biotic stress in Solanaceae crops. We tested several architectures for simultaneous classification and automated severity assessment of diseases in tomato, potato, bell pepper, and eggplant. A key feature of this framework is its ability to perform core diagnostic tasks: classification, segmentation, and severity measurement, without requiring an Internet connection, making it accessible in remote agricultural areas. The selected models were integrated into a native Kotlin mobile platform featuring an expert diagnostic module powered by Groq’s LPU (Language Processing Unit) inference engine to provide real-time analysis. Model performance was thoroughly evaluated using accuracy, recall, F1-score, Mean IoU, mPA, inference latency, and model weight. The top-performing model was DeepLabV3+ with MobileNetV2, which achieved over 97% classification accuracy while operating at lower inference latency. In addition to individual diagnostics, the system includes a collaborative georeferencing feature that allows users to share observations and precise geographic coordinates of detected diseases to support regional epidemiological tracking.