Automating exam integrity at scale

A GovTech business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, all centred on making integrity checks reliable and scalable without adding manual effort.

The pipeline is serverless and event-driven, so it scales with demand and there is no infrastructure to manage between sittings. A client requests a pre-signed upload URL from Amazon API Gateway and uploads all of a candidate's media, images and the room-scan video, into Amazon S3 under a single test identifier. Each uploaded file raises an event that Amazon EventBridge routes to a Router Lambda function, which inspects the file type and sends it down the right path.

Cloud Combinator delivered a working minimum viable product against the agreed scope and success criteria, proving that automated integrity checks are accurate enough to sit at the front of the client's proctoring process. The pipeline detects people and prohibited devices across both images and video, applies the client's violation rules consistently, and produces a clear report for every sitting.

• Automated triage — Clean sittings pass automatically while suspicious ones are flagged, so proctors spend their time only on the cases that need a human eye. A reduction of at least half the manual review workload is the target the pipeline is designed to meet.
• Accurate detection — Rekognition label detection identifies people and prohibited devices such as phones, laptops, tablets and monitors, distinguishing genuine violations from acceptable items like a household television.
• Consistent, tunable rules — A defined rule set covers room scans, in-exam photos, image quality and unusual audio-submission patterns, with a confidence threshold that can be calibrated as real-world results come in.
• Auditable reporting — Every sitting produces a structured JSON violation report in S3, giving the client a consistent, reviewable record behind each pass, fail or review decision.
• Handover — The solution was built to deploy into the client's own AWS environment, with the architecture, decision logic and reporting schema documented so their team can operate and extend it.

With the MVP proven, the natural next steps are to calibrate thresholds against production data, and, where out-of-the-box detection needs sharpening for specific edge cases, to train an Amazon Rekognition Custom Labels model on the client's own labelled examples. Further phases could add face liveness and same-person verification, extending the platform from a proctoring accelerator into a broader test-integrity capability that grows with the business.