<p>Sarcasm detection in low-resource languages such as Hinglish is challenging due to the blending of multiple languages within a single sentence and the presence of small, highly imbalanced datasets. To overcome the challenges, this paper presents two techniques to handle the imbalanced code-mixed dataset: (i) contextual augmentation techniques and (ii) focal loss with threshold tuning with multilingual bidirectional encoder representations from transformers—recurrent neural network (mBERT-RNN) models. It critically evaluates the contextual multilingual (BERT and GPT-2) augmentation techniques to handle the imbalanced code-mixed text, along with the focal loss optimization method with threshold tuning as an alternative technique. The experiments demonstrate that the BERT-multilingual augmented dataset, trained using mBERT-GRU, achieves the highest F1-score on Hinglish sarcasm. However, due to their complexity and cultural nuances, the augmentation technique frequently adds noise to code-mixed texts. Focal loss with threshold tuning improved all mBERT-RNN models, and mBERT-BiLSTM achieved the highest macro F1 at a 0.55 threshold. We establish baselines by comparing traditional machine learning and deep learning models using TF-IDF, GloVe, FastText, and mBERT and evaluate the impact of bilingual-to-monolingual translation. Additional benchmarking with transformer models (mBERT, MuRIL) in zero-shot and fine-tuned settings, together with comparisons to existing studies, provides broader performance insights. To enhance interpretability, we apply SHAP to examine word-level contributions and understand model decision boundaries. Overall, our findings highlight the challenges of imbalanced, low-resource code-mixed datasets and demonstrate the effectiveness of combining robust training strategies with explainability for accurate sarcasm detection.</p>