The timely and accurate diagnosis of DR is critical to avoid loss of vision and to improve the patient outcomes. Although some recent deep learning approaches have shown promising results in retinal image analysis, limitations include a lack of dataset diversity, absence of explainability, and difficulty in deploying the models into daily clinical practice. This paper presents a novel deep learning-based automatic diagnosis system for early diabetes identification from retinal images. The approach is facilitated and its generalizability ensured over varying imaging conditions and subject populations by the use of advanced data augmentation, multi-source data fusion, and extensive pre-processing. Through the use of explainable AI approaches, the system offers not only a state-of-the-art diagnostic performance, but also offers transparent, visual justifications for its predictions, which are crucial to building clinician trust and understanding. Comprehensive validation shows a high sensitivity and specificity, while the architecture optimization ensures real-time inference for clinical and telemedicine scenarios. The workflow, which is automated, helps to decrease the demand on ophthalmologists for screening, making it possible to deploy scalable diabetic retinopathy screening programs, which is especially advantageous for rural and underserved areas. This project fills the void between research innovation and clinical practice by providing an accessible and reliable approach to early detection and long-term management of diabetic eye disease.