DoorDash replaced expensive LLM-only moderation with a hybrid architecture that filters obvious cases using fast internal models before escalating to LLMs for nuanced scoring. The system leverages no-code workflows and backtesting to manage content-agnostic safety at scale. This approach successfully reduced safety incidents while handling millions of daily messages in real-time.
- Hybrid pipelines cut costs by filtering obvious cases with fast models before LLM processing
- Multi-axis LLM scoring handles nuanced moderation decisions where fast models fail
- No-code workflows simplify operational management and rapid iteration of safety rules
- Backtesting capabilities allow validation of policy changes before full deployment
- Architecture scales to millions of daily messages while maintaining low latency
