This paper presents a real time, multi modal security system by combining hand gesture recognition and weapon detection for the threat detections and sending alerts. The system uses a MediaPipe for 3D hand landmark detection and also light Multilayer Perceptron (MLP) to recognize emergency gestures like “Open Palm” and “Fist” with an accuracy of 96.3%. Also uses a YOLOv5s model to identify weapons with a mean Average Precision (mAP@0.5) of 91.7%, with 0.89 precision and 0.93 recall with an inference latency of 19 ms per frame. To improve robustness, a sliding window fusion module is used to smooth out gesture predictions, lowering false positives by up to a 9%. In contrast to most of the previous systems that analyze gestures and threats individually, this system proposes a single decision model is designed for a low powered deployment. When threat is detected or emergency gesture is observed, it sends a real time alert on telegram consisting timestamps and screen snaps of the incident. The system is running on a Raspberry PI 4B model which is at ~130 milliseconds per frame (7–8 FPS) which gives real time performance in every environment, also tested under various conditions such as different lighting and environmental noise to test its strength and deployment conditions. These results indicate that the framework which is proposed is very accurate, efficient and also very practical for real time deployment in places in need of security such as schools, offices, and public transport hubs.