Using AI to Accelerate Traditional ML Pipelines at Duolingo

How do agents support the traditional ML development cycle? From exploratory analysis, model protyping, to large-scale deployment & visibility, agentic workflows can accelerate the deployment of traditional ML pipelines with smaller teams. This talk discusses recent work at Duolingo using agent-supported development to test and launch an entirely new set of ML models and infrastructure over the past few months. It will outline our own specific best-practices and pitfalls as well as explore the usefulness of common analogies for AI/developer relationships that frame these practices.