Trust & Safety at Twitch is uniquely challenging, as the vast majority of content and chat interactions unfold in real time, across a wide variety of communities with different needs, cultures, and audiences. Mitigating and preventing harm means creating fast-acting models that react quickly to bad actors’ new attack vectors, while giving Creators control over their communities. We need fast-acting models to prevent harm early and scale to all of our traffic, which can be thousands of Requests per Second. In this talk we highlight one specific challenge—channel-level ban evasion—and discuss how we address ML modeling and engineering challenges and fight this behavior on our service. We will touch on a number of strategies, including the Twitch ML Infra & ML Ops ecosystem, and how we build a complete ML stack from pipelining through real-time features, model serving, feature store and other systems.