Diabetic Retinopathy (DR) is a common complication of diabetes that can cause vision loss or blindness if not detected on time. Detection of the level of DR is key to treatment and management. In this study, we introduce a hybrid deep learning model combining Vision Transformer (ViT) and EfficientNetB0, a Convolutional Neural Network (CNN) architectures for automated DR level detection in retinal fundus images. CNN is used to extract local features from the retina images and ViT captures global dependencies which helps the model to capture the pattern and abnormalities present in the retinal structures. This fusion approach combines the benefits of both and is more accurate and robust than traditional methods. Our model was trained on the APTOS BLINDNESS DETECTION dataset and evaluated with various metrics and found to be effective in detecting DR levels with high precision. An interactive web application was also developed where users can upload retinal images and get to know their diabetic retinopathy severity. This application is to help healthcare providers and patients to detect and monitor early and manage the disease proactively.