You already know: classification with imbalanced classes and threshold choice (D3); the fraud case in this bank (adversaries, delayed labels, rules plus models); event streaming and feature stores (D16, D11). This case adds spam at platform scale: hundreds of millions of messages a day, two very different costs of error, adversaries who adapt within hours, and labels shaped by the filter itself. The design questions: which actions to take at which confidence, how to stay fast where users wait, how to catch campaigns instead of single messages, and how to keep labels honest.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "train a text classifier, block what it flags". A staff candidate asks first what each mistake costs (a legitimate message blocked versus spam delivered), which surfaces need an inline decision (direct messages are seen at once; posts can be checked just after publishing), and how the adversary behaves (campaigns of near-identical messages from many new accounts). The answers turn one classifier into a system of tiered actions, a cascade of checks, campaign detection and a label pipeline.
Assumptions you state: