ml-pipeline_skill

Overview

This skill helps design and deliver production-grade ML pipelines, from feature engineering through model deployment. It focuses on reproducible, observable workflows, experiment tracking, and automated lifecycle management to move models safely into production. Use it to standardize orchestration, versioning, and monitoring for ML workloads.

How this skill works

I inspect your project goals and data flow to propose a pipeline architecture, then produce concrete artifacts: orchestration DAGs (Kubeflow/Airflow/Prefect), feature engineering code with validation, training scripts instrumented with experiment tracking, and deployment manifests. I enforce constraints like versioned data and artifacts, reproducible environments, containerized training, and comprehensive logging and alerts.

When to use it

• Building feature pipelines and a feature store for production use
• Orchestrating distributed training or scheduled retraining workflows
• Implementing experiment tracking and a model registry
• Automating hyperparameter tuning and training resource allocation
• Creating reproducible training environments and CI for ML

Best practices

• Version data, code, and models explicitly; store artifacts in versioned object storage
• Pin dependencies, fix random seeds, and use containerized training jobs for reproducibility
• Log all hyperparameters, metrics, and artifacts to an experiment tracking system
• Validate data before training with schema checks and distribution monitoring
• Separate training and inference code, avoid hardcoded credentials or hyperparameters
• Implement retries, error handling, monitoring, and pipeline alerts

Example use cases

• Design a Kubeflow Pipelines DAG that ingests raw data, computes features in a Feast-backed store, and triggers distributed model training
• Create an Airflow DAG that runs nightly data validation, feature updates, and scheduled retraining with automatic model promotion
• Instrument training scripts to log experiments to MLflow or Weights & Biases and register validated models to a model registry
• Set up automated hyperparameter tuning with Ray Tune or native cloud services and pipeline integration for promotion
• Implement CI for ML components, including unit tests for transforms, integration tests for orchestration, and deployment manifests for Kubernetes