kRAIG: A Natural Language-Driven Agent for Automated DataOps Pipeline Generation

Modern enterprise machine learning and analytics systems rely on extract–load– transform (ELT) pipelines that remain costly and expertise-intensive to build. We introduce kRAIG, an agentic DataOps system that translates natural-language specifications into executable Kubeflow Pipelines, and ReQuesAct (Reason + Question + Act, pronounced request), a human-agent interaction framework that clarifies underspecified intent before pipeline synthesis. On the complete 100-task ELT-Bench suite, kRAIG achieves 91.0% extraction-and-loading success (SRDEL) and 21.2% transformation success (SRDT), compared with 54.0% and 13.3% for the strongest of our same-model Spider-Agent and SWE-Agent reruns. These gains of 37.0 and 7.9 percentage points come at an estimated per-task cost approximately six times lower. In an end-to-end single-shot ablation that provides the same information upfront but closes the clarification channel, performance falls to 74.0% SRDEL and 13.8% SRDT. A question-quality audit further shows that kRAIG targets empirically difficult columns and that its leading guess would have matched zero rows in 4 of 7 audited value-encoding questions.