An optimized attention-based BiLSTM model for classifying harmful comments is presented in this research. It incorporates real-time data analysis and a live comment stream. Conventional machine learning models, such as CNNs and LSTMs, frequently have trouble identifying subtle types of toxicity including context - dependent hate speech, irony, and sarcasm. Traditional machine learning models, including CNNs and LSTMs fail to recognize subtle forms of toxicity, such as context-dependent hate speech, irony, and sarcasm. A BiLSTM network with an attention mechanism will be used in this proposed system in order to track the relationships of words and pick the important indicators of toxicity. A live comment feed enables real-time classification of comments and immediate feedback and moderation. More detailed analytics module on comment data allows users to monitor trends and patterns around toxicity in comments, user behavior, and comment patterns for various languages. The system produces visual reports with analysis over time of comment activity, sentiment distribution, and common toxic phrases. Automation in toxicity detection and provision of deep analysis enable the system to lead towards safer online environments while increasing moderation efficiency and scale.