AI Cheating Detector

Remote and on-site exams are easy to cheat on, and manual invigilation doesn't scale — one person can't reliably watch dozens of live video feeds at once.

• ▸YOLOv8 + face recognition detect phones, head-turns, and unknown people in frame
• ▸Multi-violation detection running in real time over a WebSocket live-video feed with on-screen overlays
• ▸Thread-based frame processing to prevent one feed from blocking another
• ▸Telegram alerts include the flagged frame for invigilator review

• ·Multi-violation detection (phone, head turn, unknown person)
• ·WebSocket live video with overlays
• ·Thread-based frame processing

Arab Open University needed exam integrity that scales: one invigilator cannot reliably watch dozens of live video feeds at once, and reviewing recordings after the fact means violations are caught too late to matter. The brief was real-time detection with evidence, not surveillance theatre.

The system is a Python service built on FastAPI. Three detection layers run over each live feed: YOLOv8 object detection catches phones in frame, MediaPipe head-pose estimation catches sustained looking-away, and face recognition flags unknown people appearing at a student's station. Video streams over WebSockets with detection overlays drawn on-screen, and frame processing is threaded so multiple simultaneous feeds stay responsive instead of queuing behind each other. When a violation fires, a Telegram bot delivers the alert with the flagged frame attached — the invigilator sees exactly what triggered it, the moment it happened.

A deliberate design decision: the system never auto-accuses. Every alert is an evidence-backed flag routed to a human invigilator who makes the call. Thresholds are tuned to that workflow — sensitive enough to catch real violations, with the human review step absorbing the edge cases a pure-automation approach would get wrong.

Invigilators receive categorized alerts with the flagged frame, so the system narrows live feeds to events that require human review.

Measurement note: this case study documents delivered functionality and architecture. Client-approved before/after KPIs, adoption data, and revenue impact are not measured or published here.