Google PM: estimating a user's sleep from smartphone sensor signals
Company: Google
Role: Product Manager
Category: Machine Learning
Difficulty: medium
Interview Round: Other
You are a product manager at Google. How would you work out when a person is asleep using the data their cellphone collects? Walk through the phone data you would use, how you would turn it into a sleep determination, and how you would check that the result is accurate.
```hint Sleep has a signature
Compare what a phone can notice about its owner during an ordinary night with what it notices during a quiet evening at home, and ask which of those differences could fool you.
```
### Clarifying Questions
- What will the sleep determination power: a sleep summary shown to the user, automatic quiet hours on the phone, or aggregated health research? Each needs a different kind of accuracy.
- Is it enough to know when the person fell asleep and woke up, or do we also need sleep quality, such as light versus deep sleep?
- Is the phone's own data the only input, or can a paired watch or other wearable be used?
- May the feature ask the user for anything, such as a usual bedtime or a quick morning confirmation?
### What a Strong Answer Covers
- A clear definition of the output (sleep period, interruptions, possibly sleep stages) tied to what it will be used for.
- The phone signals that carry information about sleep, and the situations in which each one misleads.
- How the signals are combined into an estimate, and where ground-truth labels come from to build and test it.
- How accuracy is measured, including which kind of error hurts the use case most.
- Privacy, consent and battery cost of collecting the data, with processing kept on the device where possible.
### Follow-up Questions
- The person puts the phone on the nightstand and then reads a paper book in bed for an hour. How does your approach avoid counting that hour as sleep?
- Two people share a bedroom and only one has the feature turned on. Which signals become unreliable, and what would you do about it?
- After launch, how would you tell whether people find the feature useful, beyond whether its estimates are accurate?
Overview: A Google product manager problem-solving question about working out when someone is asleep from the data their phone collects. It tests defining the output for a real use case, choosing and combining phone signals, sourcing ground-truth labels, measuring accuracy, and handling privacy and battery limits.