AI & Technical question

Build a system to predict if something can go wrong in a cab ride.

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Tests PM-level thinking about an ML safety system: defining the prediction target, signals, and how model output translates into real-time action.

Define the target precisely: predict elevated risk of a safety incident during an in-progress ride, not just post-hoc analysis.
Identify input signals: GPS deviation from the expected route, unusual stops or speed patterns, sudden cancellation attempts, and driver or rider rating history.
Propose the model approach: a real-time anomaly-detection model scoring live trip telemetry against the rider's expected route and historical norms.
Define the action layer: a risk score above threshold triggers an automated check-in, with escalation to share live location with an emergency contact or support.
Address false positives: an ordinary wrong turn from traffic must not trigger alarming interventions, so calibrate thresholds using multiple combined signals.
Define success: reduction in unresolved deviation incidents, response time from detection to check-in, and false-positive rate on the trigger.

Defines the prediction target concretely, an in-trip risk score, instead of a vague predict-problems framing.
Names specific, realistic signals, route deviation, unusual stops, cancellation attempts, grounded in how rideshare safety systems work.
Designs a graduated action layer, check-in then escalation, rather than jumping straight to alarms.
Explicitly addresses false-positive risk, since over-alerting on ordinary traffic detours would erode user trust.

Describing only a vague AI model with no concrete signals or real-time architecture.
Ignoring false positives, which would make the safety feature more annoying than useful.

Chapters of the AI PM course that teach what this question tests.