You are the founding ML engineer at a streaming service with 40 million monthly active users and a catalogue of 120,000 titles. The home screen is currently a hand-curated set of editorial rows, and engagement has been flat for three quarters.
Product wants a personalised home screen: for every user, generate 10 rows of 20 titles each, ranked for that specific user, refreshed whenever they open the app. Leadership's stated goal is a lift in "completed watch sessions per user per week", not click-through: they are explicit that they do not want a system that optimises for clickbait thumbnails.
You have a clickstream of impressions, plays, pauses, completions and abandons; roughly 2 billion events per month. You also have title metadata (genre, cast, synopsis, runtime, release year) and coarse user attributes (signup date, country, device). About 15% of the catalogue is new each quarter, and 8% of daily active users signed up in the last 7 days.
Design the end-to-end system: what runs offline, what runs at request time, how you train it, and how you know it worked.
Build the architecture on a canvas: place the components, configure them, connect them into a data flow, and write a short reason for each one. The AI reviewer grades your design against a rubric written specifically for this problem.
Your ranker is trained on data your ranker produced. How do you prevent that feedback loop from collapsing catalogue diversity?
A title is licensed in 12 of 40 countries. Where in the pipeline do you enforce that, and what does enforcing it too late cost you?
Offline NDCG improves 4% but the A/B test shows no lift in completed sessions. What are the three most likely explanations?
Minimum 6 components · needs a wide desktop screen