A broad spread of surveillance systems means a swift need to find and name acts that break rules. This improves public safety. Our work offers a two step plan for smart, instant finding of rule breaking in surveillance videos - it puts great weight on using little computer power. In the first step, a changed Video Vision Transformer (ViViT) model looks at every tenth frame from a video - this cuts the need for computer power by a large amount. On the UCF-Crime dataset, it identifies if an act is normal or not with 85% truth. It does better than older designs like CNN-LSTM and 3D-CNN. The second step uses Video Masked Autoencoders (VideoMAE) to sort rule breaking into types - it names acts such as damage, harm along with theft with 86 % truth. To support real-time monitoring and accessibility, a feature-rich web application has been developed, offering role-based access control, heatmaps, incident frequency graphs, searchable anomaly archives, live video streaming, automatic clipping of anomalous segments, bulk video uploads, and real-time threat alerts. And this system delivers a scalable and efficient surveillance solution by combining advanced AI models with a user-friendly interface. And next Future work will focus on evaluating performance across diverse datasets, validating real-world applicability, and optimizing deployment for low-resource edge devices.