Depression is a major global mental health challenge that is usually undiagnosed and therefore needs scalable and effective methods for diagnosis. This paper presents a new multi-modal hybrid model for early depression detection by combining data from social media platforms (Twitter and Reddit) and clinical datasets (DAIC-WOZ). Our approach is different from the existing text-only methods where, with the use of cross-attention mechanisms, it enables the model to give precedence to specific modality information to enhance the prediction accuracy. All the elements of the model have been powered by the latest algorithms in deep learning, including BERT, text embedding; ResNet50 to extract features from pictures; GRU processes the audio; and MLP computes insight from behaviors. Rigorous evaluation of the MADD-Net has yielded an accuracy of 89.6%, which is a humongous increment in comparison with single-modality models. High precision, recall, and ROC–AUC scores also point out the robustness of the model in identifying minor depressed states. The work addresses an issue related to the problem of dealing with data imbalance as well as multi-modal fusion that has futures in scalable public health monitoring and clinical diagnostics. Future work includes integrating video data, personalization with reinforcement learning-based predictions, and further clinical validation of a more extensive nature.