Design a personalized ad copy pipeline

WHAT THIS TESTS The interviewer wants to see systems thinking across the full lifecycle of personalization: where data comes from, how it flows, where decisions are made, and how output reaches the user. They are probing your judgment on batch versus real-time tradeoffs and operational concerns like privacy and freshness.

A GOOD ANSWER COVERS Start with data sources: clickstream and page-view events, search queries, purchase or CRM history, and a product or offer catalog. Describe ingestion using an event stream such as Kafka for behavioral signals plus periodic batch loads for catalog and CRM. Build user profiles or segments, computing slow features (lifetime affinities) in batch and fast features (current session intent) in a streaming layer. For copy generation, use a template engine seeded with profile attributes, or an LLM with guardrails for tone and brand safety. Deliver through the ad platform's API, precomputing copy variants for common segments and caching them, with a generic fallback when no profile exists.

COMMON WRONG ANSWERS Proposing one giant offline batch job that cannot react to live intent. Ignoring consent, opt-outs, and PII handling. Treating copy generation as the whole problem while skipping ingestion and delivery. Assuming infinite latency budget at ad-serve time.

LIKELY FOLLOW-UPS How do you keep latency low at serve time? Precompute and cache. How do you measure success? A/B test variants on click-through and conversion. How do you handle cold-start users? Fall back to context or popularity. How do you respect privacy regulations like GDPR?

ONE CONCRETE EXAMPLE A user browses running shoes three times this week. Streaming features flag high category intent; a batch job knows they favor a value brand. At ad-serve, the system picks a precomputed template, fills it with the brand and a discount, and returns it to the ad platform within the latency budget, falling back to a generic shoe ad if the profile is stale.