Design a Real-Time Personalization Engine

Problem Formulation#

Most personalization systems fail not because the model is wrong, but because the team never agreed on what "right" means. Before you write a single line of training code, you need to nail down what the model is actually predicting, what signals it learns from, and what latency envelope it has to live inside.

Start by asking the interviewer: what are we personalizing? Feed items, ads, search results, and product recommendations are all ranking problems at heart, but they have very different label distributions, latency requirements, and business constraints. A feed ranking system optimizing for dwell time is a different beast from an ad auction optimizing for expected revenue. Get this on the table early.