ML Pipeline Expert guides Claude through ML infrastructure in six steps: design the pipeline architecture, validate the data schema, build feature engineering, orchestrate training, track experiments and validate and deploy. Training stops and reports when data validation fails, and a model needs recorded validation metrics before deployment.
Five reference files cover feature engineering, training pipelines, experiment tracking, pipeline orchestration and model validation.
What You Get
- Pipeline definitions as Kubeflow, Airflow or Prefect DAGs
- Feature engineering code with Feast feature views and Great Expectations checks
- Training scripts with MLflow or Weights and Biases logging, plus distributed training and hyperparameter search with Optuna or Ray Tune
- Model evaluation code with explicit pass and fail thresholds and segment-level checks
- Deployment configuration with A/B or shadow rollout and a rollback strategy
Common Use Cases
- Setting up experiment tracking and a model registry for a team
- Turning a notebook into an orchestrated training DAG
- Defining a Feast feature store for training and online serving
- Adding data validation and drift checks before each training run
- Gating model promotion on evaluation metrics and a shadow deployment
What Makes It Unique
It covers the whole model lifecycle in one skill, from feature store schema to shadow deployment. The workflow halts on a failed data check and refuses to deploy without recorded validation metrics, and the orchestration reference shows the same evaluation gate in all three of Kubeflow, Airflow and Prefect.
Under The Hood
The skill is a SKILL.md with five reference files and no scripts, and the references load by topic. It writes pipeline, training and configuration code and the agent may run it, which can start training jobs, install Python packages and use local compute. The examples connect to an MLflow tracking server and a Kubeflow host over HTTP, log to Weights and Biases, read training data from files and object storage, include a launch script for distributed training, and pull credentials from environment variables. The skill does not tell the agent to fetch web pages.