Auto ML

AutoML accepts a CSV or Trino table, uses AI-assisted agents to plan and execute training, records workflow state in PostgreSQL, stores data and generated artifacts in object storage, and logs runs to MLflow.

Current execution runs inside the API pod that accepted the request. Oqullus polls status and can recover the selected active ID only from the same browser; it does not expose the backend cancellation control. Trino reads use an AutoML service account rather than the signed-in user's Ranger identity.

AutoML produces candidate artifacts and MLflow runs. It does not currently provide a consistent user-visible transition to Experiments, Models, or serving.

Task suggestions profile an uploaded CSV but do not prove execution capacity. Engine selection is automatic: internal size tiers select algorithms, while the deployed Spark runtime remains local to the API pod. Synthetic Data is a related augmentation flow with separate status/history; its output contains both source and generated rows and is not currently downloadable through Oqullus.

The backend currently provides its own AutoML engine using fixed candidate sets, small hard-coded GridSearchCV spaces, and custom preprocessing and task heuristics. MLflow tracks the resulting lifecycle state; it is not the search engine. The assessment does not preselect a replacement. See ARCH-ML-001 for a comparison of AutoGluon, FLAML, H2O AutoML, PyCaret, Optuna, Ray Tune, and Spark ML.

See Use AI/ML, the AI/ML assessment, architecture, and evidence.