Diabetic Retinopathy (DR) is a leading cause of preventable blindness, particularly in low-resource settings where regular screening is limited. While deep learning shows promise for automated DR detection, existing methods often lack generalizability, interpretability, and integration of clinical context. This study introduces a novel hybrid AI framework to address these gaps. For retinal image analysis, a meta-learning strategy is employed, where a Multi-Layer Perceptron (MLP) meta-learner aggregates high-level visual features extracted from five state-of-the-art Convolutional Neural Networks (ResNet-50, Inception-V3, EfficientNet-B0, DenseNet-121, and Xception). This image-based component is trained and evaluated on APTOS 2019 and Messidor-2 datasets. Crucially, we augment this with a separate clinical data processing pipeline. A Random Forest classifier is trained on tabular patient metadata from WiDS Datathon 2021, Diabetes Prediction, and Diabetes datasets to capture systemic risk factors. This multimodal integration aims to create a more comprehensive diagnostic system. Our framework achieves over 90% accuracy and demonstrates strong generalization across DR stages. Future work will focus on integrating Explainable AI (XAI) techniques like Grad-CAM for enhanced clinical trust and deploying the system via platforms like Telegram bots to support interactive screening in remote communities.