This work proposes a new technique for real-time violence detection in video feeds using deep learning combined with Explainable AI (XAI) methodologies. Models used are Convolutional Neural Networks (CNNs) for feature extraction as they excel at capturing spatial patterns, such as edges, shapes, and texture, making them ideal for classification with (LSTM) networks for temporal analysis as they are designed to process sequential data, capturing dependencies across time and understanding the progression of events ideal for video anomaly detection. Grad-CAM visualization highlights important regions influencing decisions, helping interpret the model and identify why specific frames are misclassified for further improvement. Integration is performed with a Telegram bot to allow the agent to continuously monitor for violence and send notifications as soon as a violent incident is observed. This also ensures ease of deployment. Its high accuracy, lightweight model, and transparency in the detection process make it perfect for deployment on cameras for surveillance or safety-centered products.