<p>This work introduces a two-stage explainable artificial intelligence framework for automated diagnosis of eye diseases. The objective is to support early screening and improve diagnostic accuracy by integrating multi-disease classification with severity grading of diabetic retinopathy. The proposed system is built on Vision Transformer models and enhanced through data augmentation, class balancing, and optimization techniques. In the first stage, the model classifies images into four categories: cataract, glaucoma, diabetic retinopathy, and normal. In the second stage, cases identified as diabetic retinopathy are further graded into early non-proliferative, severe non-proliferative, and proliferative stages. The performance of the presented architecture is evaluated using accuracy, precision, sensitivity, and F1-score. The results of the first stage shows that the Swin Transformer achieves an accuracy of 94.15%, with precision of 0.9415, sensitivity (recall) of 0.9404, and an F1-score of 0.9405. In the second stage, the Vision Transformer provides superior performance with an accuracy of 96.59%, precision of 0.9661, sensitivity of 0.9659, and an F1-score of 0.9658. In addition, the results of the ablation studies confirm that training strategies significantly influences the performance of the presented model. In order to increase transparency, clinical reliability, and interpretability, visualization techniques based on the attention mechanism are employed, which provide meaningful insight into the decision-making areas of the model. The results demonstrate that the proposed framework accurately classify multiple eye diseases and reliably assess the severity of diabetic retinopathy. The proposed approach has the potential to enhance early detection, reduce diagnostic errors, and increase trust in artificial intelligence-based tools in ophthalmology. </p>