The rapid expansion of multilingual communication on social media has led to a surge in abusive, toxic, and offensive user-generated content. Traditional rule-based filters and monolingual detection systems fail to address modern linguistic challenges such as code-mixing (English–Hindi– Hinglish), transliteration variance, and obfuscation (e.g., “id!ot”, “b!tch”). To address these limitations, this work proposes a real-time multilingual abusive comment detection framework built using a LoRA-based parameter-efficient fine-tuning approach on Multi-lingual BERT (mBERT), augmented with a custom preprocessing pipeline and prompt-based linguistic normalization. The proposed system integrates leetspeak decoding, repeated-character normalization, slang expansion, context-driven abusive-pattern recognition, and Hindi transliteration handling, significantly improving classification robustness. The fine-tuning utilizes Low-Rank Adaptation (LoRA) to enable efficient domain adaptation without full-model training costs. A Flask-based REST API and Web Interface provide real-time detection capabilities with confidence scoring and content-restriction logic. Experiments show improved F1 scores on code-mixed and obfuscated datasets, demonstrating substantial gains over baseline mBERT and rule-based systems. Future work aims to extend the system to multimodal toxicity detection, emoji-semantic embeddings, and adversarial robustness.