What problem does it solve? Teams running one-off training scripts lack repeatable, automated ML pipelines, leading to stale models, bad-data incidents reaching production, and uncontrolled GPU costs on Kubernetes. ## Core Features & Use Cases - Pipeline Step Design: Instantiate a full step taxonomy (data validation, feature engineering, training, evaluation, conditional registration) with typed inputs, outputs, and quality gates. - Orchestrator & Trigger Selection: Choose between Argo Workflows, Kubeflow Pipelines, Airflow, or TFX using explicit decision rules, and pick from four trigger types (GitOps commit, new-data event, schedule, statistical drift). - Cost Engineering: Apply step caching keyed on dataset versions and spot/preemptible GPU training with mandatory checkpoint-and-resume logic. - Use Case: A research team on managed Kubernetes wants retraining to fire automatically when new data lands in S3-compatible storage; the Skill produces a pipeline design doc with a trigger spec, caching plan, and a skeleton Argo WorkflowTemplate. ## Quick Start Use the ml-pipeline-architect skill to design an automated retraining pipeline on Argo Workflows that triggers on new data in our object store and registers models in MLflow only above a metric threshold.
