Project Panopticon — AI-Powered Exam Proctoring & Risk Analysis System
Panopticon is an AI-powered exam monitoring and risk-analysis system that combines multimodel ML, behavioral telemetry, rule-based risk scoring, and a monitoring dashboard to identify potentially suspicious exam activity.
Business Challenge
Online examinations need reliable ways to identify potentially suspicious activity while handling multiple behavioral signals such as tab switching, copy/paste activity, gaze deviation, missing-face events, and audio activity. A single signal can be noisy, so Panopticon was designed to combine telemetry, machine-learning predictions, and rule-based risk analysis into a structured monitoring workflow.
Solution Delivered
Built Panopticon as an AI-assisted exam monitoring and risk-analysis prototype. The system includes synthetic telemetry generation, feature engineering, multiple classification models including Random Forest, XGBoost, LightGBM, and CatBoost, model evaluation across decision thresholds, threshold optimization with a configurable 90% minimum-precision target, rule-based risk scoring, and a Streamlit monitoring dashboard. The system also includes separate modules for CV, audio, data processing, ML evaluation, explainability, and automated testing.
Outcomes Achieved
Panopticon demonstrates an end-to-end AI-assisted exam monitoring and risk-analysis workflow. It brings multiple telemetry signals into a unified analysis pipeline, evaluates several machine-learning models using precision, recall, F1, ROC-AUC, PR-AUC and confusion-matrix metrics, and provides configurable threshold optimization and rule-based risk categorization. The project also provides a monitoring dashboard for reviewing risk events and model analysis.
Measurable Business Outcome
The system provides measurable model-evaluation outputs including precision, recall, F1-score, ROC-AUC, PR-AUC, false positives, false negatives, and confusion-matrix results across multiple decision thresholds. It also supports threshold optimization against a configurable minimum precision target of 90%. These measurements are project-level evaluation results rather than client or commercial business metrics.