Skip to main content
Remote technical interviews are highly susceptible to generative AI cheating. Candidates frequently use tools like ChatGPT, Claude, or GitHub Copilot on secondary screens to generate live answers. Kovi utilizes a multi-layered proctoring engine to actively detect LLM assistance and hardware circumvention.

​
1. AI Copilot & Cadence Detection

Traditional proctoring relies entirely on screen sharing, which is easily bypassed by using a second device. Kovi focuses on conversational telemetry.
  • Rhythm Analysis: When candidates read LLM-generated text, their speech cadence flattens, and pauses align abnormally with the generation speed of the AI tool. Kovi monitors this audio buffer in real-time.
  • Dynamic Curveballs: If the LangGraph supervisor detects a high probability of scripted reading, it instantly injects a highly specific, architectural follow-up question. (e.g., “You mentioned using a Redis cache there. How would you handle cache stampedes in that exact microservice?”).
  • Candidates relying on an LLM typically freeze or display massive latency spikes when forced to defend a specific design choice on the spot.

​
2. Hardware & Browser Proctoring

The WebRTC client actively monitors the candidate’s physical and digital environment using standard browser APIs, requiring no external downloads.
  • Tab-Switch Logging: Captures exact timestamps and durations whenever the candidate navigates away from the active interview tab.
  • Gaze Estimation: Uses the webcam feed to track sustained eye movement away from the primary monitor.
  • Multiple Face Detection: Flags instances where a second person enters the webcam frame to assist with the interview.

​
3. The Non-Blocking Evidence Policy

Kovi does not abruptly terminate the interview if cheating is detected. Abrupt terminations often lead to aggressive candidate disputes and negative Glassdoor reviews claiming “technical glitches.” Instead, Kovi follows a non-blocking evidence collection protocol:
  1. The interview proceeds naturally to completion.
  2. All telemetry (tab switches, gaze flags, Copilot cadence alerts) is silently logged.
  3. Upon completion, the final JSON scorecard delivered to the HR team includes a proctoring_flags object.
This empowers your engineering managers to reject the candidate with concrete evidence, rather than relying on an automated system shutdown.