Malpractice during offline pen-and-paper examinations remains a persistent and largely unsolved problem because a small number of human invigilators cannot maintain consistent attention across a crowded hall for several hours at a stretch. This paper presents the design, implementation, and experimental evaluation of an AI-based, fully offline examination monitoring system that uses computer vision to detect suspicious student behaviour in real time. The proposed pipeline combines YOLOv8n for person detection, DeepSORT for persistent multi-object tracking, and MediaPipe for facial, postural, and hand-landmark extraction, feeding a lightweight rule-based behavioural-scoring engine that accumulates a suspicion score per student. A distinguishing contribution is a multi-angle supervisor-registration module, built on K-Means clustering over landmark features captured from four viewing angles, which allows the system to reliably recognise the invigilator and exclude their own movement from malpractice checks. On real classroom video, the system sustained an average of 18.7 frames per second on commodity laptop hardware, generated alerts with a mean latency of 5.39 seconds, and achieved 93.17% head-pose classification accuracy, 96.0% supervisor-identification accuracy, and 94.0% accuracy for combined cheating and passing detection, entirely without internet connectivity. These results indicate that a carefully engineered combination of existing open-source computer-vision components can deliver a practical, affordable, and privacy-respecting alternative to manual-only invigilation