Diabetic Retinopathy (DR) is still a significant cause of avoidable blindness that needs early diagnosis to be treated clinically. Despite the popularity of deep learning, interpretable and computationally efficient Artificial Intelligence models based on handcrafted features can be highly valuable in massive screening. However, the presence of redundant features, inconsistent image quality, and overlapping disease stages often limits the performance of traditional classifiers. To address these challenges, this study proposes an optimized AI-based classification framework using handcrafted feature extraction and ensemble learning to identify early and reliable DR. A total of 22,454 fundus images were preprocessed (resizing, CLAHE enhancing, and normalization), and then, 28 features (intensity, texture, edge, wavelet, and illumination) were extracted. The feature selection using correlation was done at varying thresholds, and at a 0.98 threshold, the optimal subset of 21 highly discriminative features was gained. Eleven machine learning models were assessed, and the best of the lot were Extra Tree and XGBoost, which were further optimized with Bayesian hyperparameter tuning. Lastly, four ensemble strategies (Hard, Soft, Stacking, Weighted) were performed to enhance predictive stability. The best performance was reached by the Soft Voting ensemble with an accuracy of 0.9578, best precision, recall balance, excellent MCC and Kappa, and AUC of 0.97, which was better than the individual models. These results indicate that handcrafted feature optimization with a high level of ensemble learning can create a strong and interpretable solution to the early DR detection, which is clinically compatible and can be used computationally efficiently.