ExamX Protocol: AI-Powered Examination Integrity System

ExamX Protocol is a comprehensive, AI-powered examination integrity framework designed to significantly strengthen trust, fairness, and accountability within online assessment environments. The system's core functionality revolves around creating a highly controlled examination setting.

We developed a custom Electron-based secure browser that enforces strict restrictions, including fullscreen mode, blocking of keyboard shortcuts, continuous focus monitoring, and restricted navigation. This controlled environment ensures that students interact solely with the examination material.

The backend operates on a modular, API-driven architecture. Key components include RedisAPI for central stream coordination of webcam and microphone data, GazeAPI for analyzing visual behavior (using face mesh, gaze direction, and head pose estimation), and WhisperAPI for transcribing spoken audio. Furthermore, an LLMAPI interprets this transcribed speech against the specific Moodle exam context to classify conversations as casual, procedural, suspicious, or related to answer assistance. Finally, the CheatingAlgoAPI synthesizes signals from visual, audio, semantic, temporal, and browser-policy layers into a unified, explainable Cheating Likelihood Score.

Crucially, the system is engineered as a decision-support tool for human reviewers, not an automated punitive system. Every generated score is accompanied by structured, auditable evidence—such as records of off-screen gaze, detected suspicious speech, or focus loss—allowing lecturers and administrators to review suspicious sessions with unparalleled transparency and accountability.

Deployment and scalability were addressed using industry standards like Docker and Kubernetes, demonstrating a robust, production-ready platform for digital and hybrid assessment needs.

The primary challenge was integrating disparate, complex systems secure browser enforcement, real-time media streaming, advanced AI analysis (visual, audio, semantic), and risk scoring—into one cohesive, non-fragmented decision-support tool. We also had to ensure the output was explainable, resisting the temptation of a black-box AI verdict, while maintaining high security against injection and misuse.

We adopted a modular, API-based architecture, using RedisAPI as a central stream coordination layer to manage real-time data flow and prevent backend overload. Each AI component (GazeAPI, WhisperAPI, LLMAPI) processes its data independently, but the CheatingAlgoAPI acts as the final fusion point, combining all evidence streams into a single, transparent score. Moodle integration provided the necessary context and workflow management, while Docker/Kubernetes ensured scalable, consistent deployment.