How DoorDash's AI moderation actually works at scale

DoorDash has built an internal system for routing millions of user messages through an LLM to catch things like fraud, scams, and policy violations. The post breaks down the architecture: it's not a single model call but a layered pipeline with different models handling different signal types, each writing its own log so the whole thing is traceable.

The key move is LogDrive, a shared logging layer that lets each stage in the moderation pipeline write structured events without knowing what the other stages are doing. This keeps the system flexible — you can swap models, add new checks, or reroute traffic — without rewriting every component. The logs become the source of truth for what actually happened, not just what the code says it did.

Why this matters for us:
When platforms use AI to moderate content, the decisions are opaque and often irreversible — building systems that actually log what the AI did, and who's accountable, is the bare minimum we should expect.