| name | automl-propensity |
| description | Train and deploy propensity models using Databricks AutoML. Use when the agent needs campaign-specific propensity scores and no Model Serving endpoint exists yet, or when retraining is needed for a new campaign type. |
AutoML Propensity Model Training
Spawn a Databricks AutoML workflow to train a propensity model for a specific campaign type (e.g., cart_recovery, seasonal_purchase, loyalty_renewal). The trained model is registered in Unity Catalog and deployed to a Model Serving endpoint.
When to Use
- No propensity model exists for a campaign type
- Existing model is stale (>30 days since last training)
- New campaign type requires its own propensity scoring
- Agent needs real-time propensity scoring instead of cached scores
What It Does
- Submits an AutoML classification experiment on customer behavior data
- Selects the best model from the experiment
- Registers the model in UC:
maestro_cdp.models.propensity_{campaign_type}
- Creates or updates a Model Serving endpoint
- Returns the endpoint name for
score_propensity tool to call
Scripts
train_propensity.py
Train a propensity model for a given campaign type.
Arguments:
--campaign-type (required): The campaign type to train for (e.g., cart_recovery)--training-table (optional): UC table with training data (default: maestro_cdp.features.customer_events)--target-column (optional): Target column name (default: converted)--profile (optional): Databricks CLI profile (default: 9cefok)
Output: JSON with model_name, endpoint_name, experiment_url, accuracy
Beat 2 vs Beat 2.5+
- Beat 2:
score_propensity returns cached synthetic scores (0.81 for Cindy). This skill is discoverable but not invoked.
- Beat 2.5+: Agent can invoke this skill to train real models. Once the endpoint is live,
score_propensity calls it instead of returning cached data.
Prerequisites
- Databricks workspace with ML Runtime
- Feature table with customer behavior data
- Unity Catalog with model registry enabled
