Python MLOps

Develop machine learning workflows from exploration to evaluated delivery, grounded in the four reviewed MLOps repositories.

Routing

Read only the matching guide. For notebook authoring or reactivity, go directly to marimo; experiment tracking and model delivery are separate tasks. Preserve the project's libraries and notebook format. The reference package illustrates boundaries, not a mandatory architecture.

For ML work, establish permitted compute and artifact destinations before running jobs. Training, tracking, registry writes, and promotion have different effects; inspect the selected task's dependencies before execution.

Task guides

• experiments: Track reproducible MLflow experiments, data lineage, metrics, model signatures, and bounded artifacts.
• marimo: Create, debug, convert, and export reactive marimo Python notebooks and apps, including agent pairing.
• ml-jobs: Turn notebook experiments into typed, configurable ML jobs; select the MLOps template or reference architecture.
• model-delivery: Evaluate exact model versions, manage MLflow registry aliases, and verify batch inference and rollback.
• polars: Transform tabular data with Polars lazy queries and expressions; hand frames to Pandera, scikit-learn, and DuckDB.
• training: Validate Polars or pandas data with Pandera, prevent leakage, and train, tune, and evaluate scikit-learn pipelines.

Related owners

• python-stack: uv, typing, packaging, and async; python-testing: test implementation.
• Colab, Kaggle, and Hugging Face retain connector ownership of accounts, sessions, transfers, and compute. Available quota does not authorize allocation.
• duckdb: SQL and file queries; python-web: APIs and model demos; observability: service telemetry; agent-evaluation: stochastic agent evaluation.