Scale AI engineers describe how they built an augmented data processing platform on top of Temporal for evaluating and improving deployed AI agents. The system uses a generic 'run card' activity pattern where each step reads data, applies a custom transformation, and writes to a dataset service — bypassing Temporal's data size limits by passing URI references between activities instead of raw data. Temporal signals enable an interactive, step-by-step workflow building experience in the UI. The architecture supports joining data across workflows, spawning child workflows for cross-language execution (TypeScript calling a Python service), and eventually running all steps programmatically in one shot.
•16m watch time
229 Impressions