Diabetic Retinopathy (DR) is a leading cause of preventable blindness globally, demanding accurate and early detection for  effective intervention. This study introduces a Randomization-Driven Hybrid Transformer Deep Learning (RD-HTDL) framework  for multi-class DR detection, combining Convolutional Neural Networks (CNNs) for local feature extraction with Vision Transformers (ViT) for global context modeling. To enhance model robustness and generalization, randomized weight initialization and advanced hyperparameter optimization using a Bayesian-Genetic hybrid search algorithm were employed. The framework was trained and validated on Messidor-2 and APTOS 2019 Blindness Detection datasets,encompassing over 10,000 high-resolution retinal fundus images collected from clinical sources in Europe and India. Each image underwent preprocessing steps including contrast-limited adaptive histogram equalization (CLAHE), normalization, and extensive data augmentation, addressing noise, illumination variance, and class imbalance. Experimental results demonstrate that the RD-HTDL model achieves accuracy, precision,  recall, and F1-scores above 96%, outperforming standalone CNN, ResNet50, DenseNet121, and standard ViT architectures in multi-class classification of DR stages (No DR, Mild, Moderate,  Severe, Proliferative). The hybrid CNN-Transformer architecture captures both fine-grained  retinal features and global spatial dependencies, while randomization-driven training ensures  fast convergence and minimized overfitting. This framework is scalable, extendable, and suitable  for integration into teleophthalmology and automated screening systems, offering a robust tool to assist clinicians in early stage DR detection and reduce manual diagnostic errors.