ML Full Cycle Orchestration

Instructions

You are executing a production-grade Machine Learning pipeline. [cite_start]When a user provides a dataset and a target variable, follow this sequential workflow orchestration[cite: 486].

Step 1: Data Validation & EDA

1. Inspect the provided dataset format (CSV, Parquet, JSON).
2. Identify the target column specified by the user.
3. Conduct brief univariate analysis to identify categorical vs. numerical features and check for severe class imbalances.

Step 2: Pipeline Execution

Run the generalized ML pipeline script to handle preprocessing, training, and tracking. Execute the following command, replacing the variables with the user’s inputs:

python scripts/generalized_ml_pipeline.py --data_path "PATH_TO_DATA" --target_col "TARGET_COLUMN_NAME" --experiment_name "EXPERIMENT_NAME"

Step 3: Observability & Evaluation

After the pipeline finishes execution:

1. Review the generated classification_report_[model].txt and confusion matrices.
2. Select the champion model based on the highest F1-Score (to account for potential class imbalances).
3. Translate the technical metrics (Precision, Recall, F1) into business impact.

📌 Reference: Consult references/business_kpi_mapping.md for business translation frameworks.

📄 Output Requirement: Format your final pipeline summary using the structures defined in assets/model_card_template.md and assets/data_storytelling_report.md.

Step 4: MLOps Deployment & Serving

Transition the champion model to the deployment phase:

1. Explain the architectural trade-offs between Shadow Testing, Canary Releases, and A/B Testing.

2. Guide the user to serve the model using the provided FastAPI script:

   python scripts/serve_model.py
   

3. Advise the user to containerize the solution for Kubernetes or Cloud Run using the scripts/Dockerfile.

📌 Reference: Consult references/deployment_strategies.md and references/monitoring_observability.md to define a monitoring strategy (tracking Inference Latency and Data Drift).

Common Issues & Troubleshooting