You already know: click and conversion prediction with calibration (the ad CTR case), budget-split experiments (the ad CTR case), control loops (the surge pricing case) and in-memory serving under tight latency (the autocomplete case). This case is the buying side of programmatic advertising: a bidder receives millions of auction requests per second from exchanges, must answer within milliseconds, values each impression for many advertisers, bids strategically in first-price auctions, and spends each advertiser's budget evenly and never more. The design questions: how to value, how much to bid, how to pace and cap budgets across thousands of servers, and how to stay safe when anything is slow.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "predict click-through rate and bid that times the advertiser's price per click". A staff candidate first asks the auction type (first-price or second-price, floors), the advertisers' goals and budgets, the volume and timeout, and the rules on traffic quality and consent. Those answers produce bid shading, value-based bids with pacing, an in-memory hot path that can always answer no-bid, and filters for invalid traffic, consent and brand safety.
Assumptions you state: