—Hinglish and other code-mixed forms of languages have experienced an increase in hate speech owing to the avail ability of user-generated content and the difficulties of automated moderation in coping with language changes, informal spelling. Prior work on this problem, including an earlier version of this study, proposed a fusion of lightweight transformer encoders with a symbolic decision layer; a rigorous re-evaluation showed that the reported symbolic-reasoning component was not implemented and that the earlier evaluation protocol produced unreliable metrics (near-random ROC-AUC despite high accuracy). This paper presents a corrected and fully re-implemented pipeline: a TF-IDF with Linear Support Vector Classifier (SVC) baseline, a fine-tuned DeBERTa transformer, and a probability-averaging ensemble of the two. The framework further incorporates In tegrated Gradients for token-level explainability and a targeted bias probe over identity-related sentences. On a held-out test set of 4,744 code-mixed Hindi-English samples, the baseline attains 88.89% accuracy, the fine-tuned DeBERTa model attains 91.82% accuracy, and the probability ensemble achieves the best overall performance at 93.39% precision, 91.99% accuracy, 92.81% recall, and ROC-AUC of 0.9715, with a 93.09% F1-score and an average precision of 0.9806. The bias probe indicates that none of the tested identity-related sentences crossed the decision threshold of 0.45, suggesting limited identity-based bias. The results demonstrate that a well-calibrated transformer ensemble, combined with post-hoc explainability and fairness auditing, provides an accurate, interpretable, and reproducible approach to Hinglish hate speech moderation.

Hate speech detection, code-mixed NLP, Hinglish, transformer ensemble, DeBERTa, explainable AI, Inte grated Gradients, bias evaluation.