Product Context

FitPulse is a consumer mobile app for workout tracking and wellness content. The team wants to add ML-powered personalization across the home screen: recommended workouts, articles, and push notification suggestions tailored to each user.

Scale

Users open the app 3.5 times/day on average, and the home screen returns 20 ranked items per session. About 40% of users are relatively cold-start each month with fewer than 5 meaningful interactions. The business goal is to improve 7-day retention and workout starts without materially increasing notification fatigue or serving cost.

Deliverables

1. Define the functional and non-functional requirements for ML features in the mobile app.
2. Propose an end-to-end architecture covering data collection, feature computation, model training, and online serving.
3. Design a multi-stage recommendation pipeline for the home screen (retrieval → ranking → re-ranking) and explain whether push notification prediction should reuse the same models or use a separate path.
4. Choose model types for each stage and justify them with latency, freshness, and interpretability tradeoffs.
5. Define offline evaluation, online experimentation, and monitoring for drift, training-serving skew, and degraded user experience.
6. Identify major failure modes, fallback behavior, and how you would launch safely.

Constraints

• Mobile clients are often on poor networks, so payload size and response time matter.
• User-level health data is sensitive; raw event retention is limited to 30 days in some regions.
• Freshness matters: newly published workout plans should become eligible within 10 minutes.
• The team has a moderate infra budget: GPU serving is allowed only for the heaviest stage, and only if clearly justified.
• Product requires guardrails for over-notification, content diversity, and safe recommendations for beginner users.