Social media plays a major role in contemporary life. Social media growth has generated enormous user data, including toxic, and abusive comments, which affect both individuals and society. Addressing this issue is of utmost importance to ensure safer digital interactions. In this study, we present a hybrid lightweight approach for Tamil-English code-mixed toxic and abusive comment classification, ensuring interpretability and transparency by incorporating multiple learning paradigms. In this paper, we propose a novel approach of combining lightweight models which achieves higher performance through efficient learning paradigm. FastText embeddings are used for feature representation, and L2-regularized Logistic Regression is applied to enhance classification performance. To further strengthen toxicity detection, we integrate Zero-Shot Learning using DistilBERT MNLI and One-Shot Learning with a fine-tuned MuRIL model. These methods enable effective toxicity classification even in resource-constrained settings. The outcome of the experimentation demonstrates the effectiveness of our approach in improving classification accuracy compared to traditional machine learning models.