How can you quickly get to a minimal viable AI system that can be iteratively improved with confidence through testing and easily operated with observability? How long should your prototype take to build - 1 hour, 1 day, 1 week, one month?If you follow MLOps guides from the dominant players, you will have >40 boxes to build before you get to a working system.In this talk, I will show how can we reduce that to 1 hour or 1 day by introducing an architecture for AI systems based around feature/training/inference pipelines and a shared state layer. We will follow the principles of modularity and composability for system design, and automated testing, versioning, and observability for systems operation. But, we will enable you or your teams to independently develop, test, and operate the pipelines that make up your AI system, connected by a feature store, model registry, and any other infrastructure you need to operate your AI system.

More info