DoorDash built a two-stage moderation system to screen more than four million chat messages exchanged each day without putting every message through a large language model. The company says an LLM-only design would have added variable delays of roughly two to ten seconds and made per-call costs impractical.
SafeChat first removes noise and sends messages to a small internal classifier trained to identify content that is obviously safe. DoorDash designed this layer to handle most traffic cheaply and return 90% of responses in under 100 milliseconds. Fewer than 10% of messages then move to the slower LLM layer for closer review.
Instead of asking the LLM for a simple safe-or-unsafe answer, the system requests scores across dimensions such as threats, profanity, and sexual content. Those scores let DoorDash apply graduated actions: censor lower-severity language, block insults, or offer cancellation and additional protections for serious threats.
The architecture followed months of data collection showing that only a small single-digit percentage of messages were unsafe. DoorDash presents the pattern as useful beyond moderation: send common cases through a fast specialized model and reserve a more capable, expensive model for ambiguous cases. Its reported results come from DoorDash's own production system and conference presentation.