The disease Diabetic Retinopathy stands as the most common origin of vision deterioration along with blindness which affects diabetic patients worldwide. Proper and early detection lays the foundation for quick professional intervention in cases of DR. Clinical screening programs from ophthalmologists depend on labor-intensive manual evaluation that limits mass screening ability. The hybrid deep learning system that combines EfficientNet-B3 with ResNet-50 and Coordinate Attention was developed to advance DR classification precision. The model makes use of sophisticated data preprocessing with CLAHE and random cropping and flipping and rotation and contrast enhancement to achieve robust performance. The feature acquisition process of EfficientNet-B3 benefits from capturing fine texture patterns and ResNet-50 provides a high-level understanding of the image content. The classification of fused features consists of No DR and five other progressively severe DR categories beginning with Mild and ending with Proliferative DR. The built model demonstrates an 83.61% accuracy evaluation together with a 0.9051 quadratic-weighted Cohen’s Kappa Score for accurate expert diagnosis assessment. The assessment of performance uses precision, recall, and F1-score in combination with confusion matrix analysis. The findings suggest that AI- driven DR screening can enhance early diagnosis and assist ophthalmologists in large-scale screening programs, particularly in resource-limited settings.