Preventing vision loss and enhancing patient outcomes need the early and precise diagnosis of ocular conditions such age-related macular degeneration, glaucoma, and diabetic retinopathy. Physical assessment is the foundation of traditional diagnostic techniques, however it can be laborious and subjective. Recent developments in deep learning have shown great promise for automating the use of fundus images to detect eye diseases. Diagnostic accuracy is affected by the limits of current models, which frequently fail to capture both local and global image features. In order to improve the precision and robustness of ocular disease categorisation, we present a unique fusion framework in this work that combines Vision Transformers (ViTs), EfficientNet, and ResNet-50. ResNet-50 retrieves hierarchical spatial characteristics, EfficientNet assures computational efficiency, and the ViT component efficiently captures long-range dependencies. We also present a risk-grading module that gives clinicians important information for decision-making by evaluating the severity of the disease. The ODIR-5 K dataset is used to test the suggested model, and it outperforms state-of-the-art methods with an accuracy of 97.8% and an AUC-ROC of 0.98. With a Mean Absolute Error of 0.12, the risk-grading module offers a quantifiable assessment of disease severity, significantly improving clinical applicability. The experimental findings demonstrate how well the fusion technique works to increase classification accuracy while preserving processing efficiency. By bridging the gap between deep learning research and practical clinical applications, our work advances AI-driven retinal disease diagnosis. Future initiatives include optimising the system for real-time deployment in clinical settings and incorporating multi-modal imaging data.