🤖DoorDash's SafeChat Filters 4M Daily Chats for Safety
DoorDash filters millions of daily chats with AI to keep users safe
TL;DR
DoorDash's SafeChat uses an ML model and LLM to filter over 4M daily chats, ensuring only safe messages reach their destination. The system identifies unsafe content with high accuracy while minimizing costs.
DoorDash has launched SafeChat, an AI-powered safety system designed to classify chat messages as safe or not in real-time. Over 4 million chat messages are exchanged daily on the platform, and SafeChat ensures that each message is classified before reaching its destination. The system uses a three-layer architecture: noise removal, a cheap ML model for quick classification of obvious safety issues, and an expensive LLM for complex cases. This approach minimizes costs while maintaining high accuracy in identifying unsafe content. Developers working on similar real-time moderation systems can learn from DoorDash's cost-effective solution.
Key Points
SafeChat processes over 4M chat messages daily on the DoorDash platform
The system classifies over 200K images exchanged in chat or SMS every day
Only a single-digit percent of messages are flagged as unsafe by SafeChat
A cheap ML model is used for quick, cost-effective classification of obvious safety issues
An expensive LLM is called only when the cheaper layer can't classify a message
Why It Matters
If you're building real-time moderation tools for chat platforms, DoorDash's SafeChat architecture offers valuable insights. The system uses a three-layer approach to balance accuracy and cost, ensuring millions of daily messages are safe without breaking the bank.
Frequently Asked Questions
Why does this matter?
If you're building real-time moderation tools for chat platforms, DoorDash's SafeChat architecture offers valuable insights. The system uses a three-layer approach to balance accuracy and cost, ensuring millions of daily messages are safe without breaking the bank.
What happened?
DoorDash's SafeChat uses an ML model and LLM to filter over 4M daily chats, ensuring only safe messages reach their destination. The system identifies unsafe content with high accuracy while minimizing costs.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,304 builders reading daily.