Presentation: SafeChat: Building AI-Powered Safety Systems at Scale in a Real-Time Marketplace
DoorDash cut safety-incident costs by hybrid-stacking: fast internal models + LLM multi-axis scoring. Here's the architecture.
Why it matters
A real production case study showing how to build scalable AI moderation without LLM-only economics. Practitioners deploying safety systems at scale will see actionable architectural patterns.
The key facts
11 to knowHybrid moderation pattern: fast internal models for obvious cases + LLM for nuanced multi-axis scoring
Replaces costly LLM-only pipelines
No-code workflows with backtesting built in
Scaled to millions of daily messages
Cut safety incidents while reducing cost
Content-agnostic design (reusable across use cases)
Hybrid architecture: fast internal models for obvious cases + LLM multi-axis scoring for nuance
No-code workflows with backtesting capability
Scales to millions of daily messages
Reduced safety incidents vs. LLM-only pipeline
Cost optimization through architectural pattern (internal models first, LLM as refinement layer)
Go to the source
InfoQ AI/MLinfoq.com
Publisher excerpt: Bruna Pereira explains how DoorDash built a content-agnostic AI moderation platform. She covers replacing costly LLM-only pipelines with a hybrid pattern: using fast internal models to filter obvious cases, LLM multi-axis scoring for nuanced decisions, and no-code workflows with backtesting.…