Post by Confluent

Real-time safety at AI scale doesn't have to mean ballooning infrastructure spend. Maor B. from Character.AI joins Joseph Morais to share how his team built real-time prompt safety testing and experimentation pipelines on WarpStream, reducing total streaming costs by 85 to 90% compared to their original approach. By running stateless WarpStream agents on already-provisioned GPU-node compute and writing directly to object storage, Character.AI unlocked: 🔷 Massive cost savings without sacrificing Kafka compatibility 🔷 A BYOC architecture that kept user data inside their own environment 🔷 A flexible foundation that grew into ETL, reverse ETL, and a custom feature store See what happens when infrastructure gets out of the way: https://cnfl.io/4vu3nlv

Video Content