In this paper, we propose a hybrid lightweight deep learning model using MobileNet and Xception to classify waste into three categories: Organic, Recyclable, and Harmful. The proposed model adopts the lightweight depthwise separable convolutions for efficiency to fit in mobile/edge devices, including smart bins and embedded IoT devices. Understanding the requirements for low-cost, high-accuracy waste sensing systems in constrained urban and ecologically sensitive spaces, the proposed solution prioritizes low computational complexity while maintaining classification accuracy. To further enhance generalized learning and reduce overfitting, the model is trained using several data augmentations, including rotation of the image, shift, zoom, and horizontal flip. This strong training strategy allows the model to deal with real-world variances, like cluttered backgrounds, presence of mixture wastes, and lighting inconsistencies. The computationally efficient model was trained and tested on a balanced and representative dataset and reached an overall test accuracy of 83%. Performance evaluations using precision, recall, F1-score, and ROC-AUC were conducted on all categories, and it was observed that the Organic and Recyclable classes were well-separable, and the Harmful class, though more difficult, achieved acceptable classification performance. By breaking down a confusion matrix and looking at class-wise performance, we see the model’s reasonable performance. The proposed system exhibits impressive potential for implementation in real-time sustainable waste management infrastructures, bringing a significative move forward for automation in ecological protection. It also opens opportunities for further work for adaptive waste sorting and on-site deployment in feedback-controlled smart environments.