TokenWave AI Pte. Ltd.
Get Start

Trustworthy Online Exams at Any Scale

Client:
TokenWave AI (In-House R&D)
Type of Work:
Computer Vision / Behavioural AI
Industry:
Education & Assessments
Stage:
Research-Based Solution
At a Glance

Case Study Overview

The Challenge

Manual video invigilation cannot scale, and malpractice undermines the credibility of online exams.

Our Solution

AI-powered proctoring with face, eye-gaze and head-pose analysis, live alerts and automated reports.

Technology

Computer vision, face detection, gaze and head-pose estimation, ProctorNet-based research.

The Outcome

Consistent monitoring for every candidate, with invigilators focusing only on flagged moments.

Executive Summary

Online examinations have become a standard part of education, certification and recruitment. They are convenient and scalable - but only if institutions can trust the results. Manual supervision over video is expensive, inconsistent and simply impossible at scale.

TokenWave AI developed an AI-powered proctoring system that monitors each candidate through their webcam, detects suspicious behaviour as it happens and produces a detailed, evidence-based report at the end of the exam. Human invigilators remain in control, but they review only the moments that matter.

Background

The Industry Context

Universities, coaching institutes, certification bodies and recruiters now conduct thousands of assessments online. Each candidate may be in a different location, using a different device, in an environment the institution cannot see or control.

Traditional online invigilation relies on staff watching live video feeds or reviewing recordings afterwards. One person can meaningfully observe only a handful of candidates at a time, and reviewing hours of footage after the exam is impractical.

project
project
The Challenge

The Problem We Set Out to Solve

1
Manual Invigilation Doesn't Scale

One invigilator cannot effectively watch dozens of candidates on video at the same time.

2
Evolving Malpractice

Looking away at notes, using another device or having someone else in the room are hard to catch consistently.

3
No Reliable Evidence

Without structured, time-stamped records, malpractice decisions are difficult to justify or defend.

4
Fairness to Honest Candidates

Weak supervision devalues results and penalises candidates who follow the rules.

Why Existing Approaches Fall Short

  • Human attention fades over long exam sessions, leading to inconsistent supervision.
  • Browser lockdown tools alone cannot detect what happens away from the screen.
  • Post-exam video review is slow and rarely done thoroughly.
  • Decisions without evidence expose institutions to disputes and appeals.

What Was at Stake

  • Credibility of results: certificates and scores lose value when malpractice goes undetected.
  • Operational cost: staffing live invigilation for large exams is expensive.
  • Institutional reputation: a single malpractice scandal can damage trust in an institution's assessments.
  • Candidate fairness: honest candidates deserve a level playing field.
Our Approach

How We Approached the Problem

We followed a structured, problem-first process - understanding the real-world problem before choosing the technology, and validating every stage before moving to the next.

Understanding Exam Workflows

We analysed how online exams are conducted, what forms of malpractice are most common and what evidence institutions need to take fair action.

Behaviour Signal Design

We defined the visual signals that indicate suspicious behaviour - absence of a face, multiple faces, sustained looking away and unusual head movement.

Model Development

Computer vision models were built for face detection, eye-gaze estimation and head-pose estimation, drawing on research including an automated proctoring system and ProctorNet.

Real-Time Alerting Logic

Rules and thresholds were designed to convert raw signals into meaningful alerts, reducing false alarms from natural, momentary movements.

Reporting and Review

Every exam produces a structured report with time-stamped events and flagged moments, so invigilators can review evidence quickly and make informed decisions.

The Solution

How the Solution Works

An automated proctoring system that monitors candidates through the webcam, detects suspicious behaviour in real time and generates an evidence-based report for every exam.

Solution Workflow

STEP 01

Candidate starts the exam with webcam on

STEP 02

Face presence is verified continuously

STEP 03

Eye gaze and head pose are tracked

STEP 04

Suspicious events trigger live alerts

STEP 05

Events are time-stamped and logged

STEP 06

Examination report is generated

Key Capabilities

Face Presence Detection

Detects when the candidate leaves the frame or when no face is visible.

Multiple Person Detection

Flags the presence of more than one face in the camera view.

Eye-Gaze and Head-Pose Tracking

Identifies sustained looking away from the screen.

Real-Time Alerts

Notifies invigilators the moment suspicious behaviour occurs.

Automated Exam Reports

Time-stamped logs and flagged events for every candidate.

Scalable Monitoring

Monitors large numbers of candidates simultaneously with consistent standards.

Technology & Techniques

PythonComputer VisionOpenCVFace DetectionEye-Gaze EstimationHead-Pose EstimationDeep Learning (ProctorNet-based research)Real-Time Alerting
Results

Business Impact & Outcomes

Credible Assessments

Consistent monitoring strengthens trust in online exam results.

Focused Invigilation

Staff review flagged moments instead of watching hours of video.

Evidence-Based Decisions

Time-stamped reports support fair and defensible outcomes.

“Online exams should be as trustworthy as those held in an exam hall. Our aim was to make fair supervision possible for every candidate, at any scale.”
— TokenWave AI Project Team

Who Can Benefit

  • Universities and colleges for semester exams and internal assessments.
  • Certification bodies for protecting the value of their credentials.
  • Recruiters and HR teams for fair online aptitude and skill tests.
  • Ed-tech platforms for adding trusted assessments to their courses.
Looking Ahead

Key Learnings & What's Next

Key Learnings

  • Combining multiple signals (face, gaze, head pose) is far more reliable than any single indicator.
  • Well-tuned thresholds are essential to avoid penalising natural movements.
  • Clear, reviewable evidence matters as much as detection itself.

What's Next

  • Audio analysis to detect speech and background voices.
  • Detection of mobile phones and other prohibited objects.
  • Integration with popular learning management systems.
  • Privacy-first processing options for on-premise deployment.

This case study describes an in-house research and development project by TokenWave AI. Outcomes are described qualitatively; actual performance depends on the data, environment and scale of each deployment. Want to apply this solution to your business? Book a free consultation.

More Case Studies

Agriculture
Crop Disease, Caught Early
 Healthcare
Explainable Medical Imaging
 Retail, Brands & Services
Hearing the Customer Clearly
 Banking & Fintech
Payment Fraud, Flagged Fast
 Cybersecurity & IT
Network Attacks, Spotted Early
 Transportation & Logistics
Driver Fatigue Alerts

Have a Process That Should Run Itself?

Tell us about it in a free, no-obligation working session. We will show you where an agent fits, what it would take and how you would measure the result.

Book a Free Session

Join our mailing list for field notes on applied AI, delivered occasionally.

Company
Services
Resources
Follow Us

Copyright © TokenWave AI Pte. Ltd. All rights reserved.