Ride-Hailing Safety: Predict Which Trips Are at Risk and Decide How to Intervene
Company: Uber
Role: Product Manager
Category: Machine Learning
Difficulty: medium
Interview Round: Other
You are a product manager at Uber. Design a system that predicts whether something is likely to go wrong on a ride. Decide what "going wrong" should include, who acts on a prediction and how, which data and modeling approach you would use, and how you would know the system is working.
```hint Define the failure before the model
"Something going wrong" covers events with very different frequency, severity and warning signs. Decide which events you will predict, who responds to a prediction and how quickly, then work back to the data and the model.
```
### Clarifying Questions
- Which failures are in scope: safety incidents such as crashes or assaults, vehicle breakdowns, route deviations and long unexplained stops, or service problems such as cancellations and fare disputes?
- Should the system predict risk before the trip, at matching time, or continuously while the trip is under way?
- Who acts on a prediction: the rider and driver through the app, a safety operations team, or automated changes such as rerouting or rematching?
- Which signals may we use, given privacy and fairness limits: phone sensor data, live GPS, past ratings and previous reports?
### What a Strong Answer Covers
- A breakdown of what can go wrong on a ride, sized by frequency and severity, with a reasoned scope for the first version.
- Clear users and actions for each level of risk, including how false alarms are handled.
- Data, label and model choices suited to rare, under-reported events that must be scored in real time.
- Offline and online evaluation, with success metrics and guardrails such as alert fatigue and the review team's workload.
- Privacy, fairness and due process for riders and drivers whose trips are flagged.
### Follow-up Questions
- Serious incidents are rare and are often reported days later. How would you build training labels and evaluate the model?
- The safety team can review only a fixed number of flagged trips per hour. How would you set the alert threshold?
- Riders in one neighborhood complain that their trips are flagged far more often than elsewhere. How would you investigate and respond?
Overview: A technical product question about designing a system that predicts when an Uber ride is likely to go wrong. It tests defining which failures matter, choosing data and labels for rare events, deciding who acts on an alert, setting thresholds, evaluating predictions, and balancing safety against privacy and fairness.