MLOps model deployment pipeline

Automated retraining pipeline with PSI-based drift detection, model versioning, and champion/challenger promotion for a churn prediction model.

Drift detection Model versioning CI/CD Champion/challenger Docker GitHub Actions

6 steps

Automated pipeline stages

Overview

End-to-end ML pipeline with data validation, training, evaluation, and conditional promotion

Production ML models degrade over time as incoming data distributions shift. This pipeline wraps a churn prediction model in six automated stages: data validation, feature engineering, model training, holdout evaluation, champion/challenger comparison, and conditional promotion. PSI monitors every numeric feature for drift, and the pipeline runs automatically via GitHub Actions on every push to main.

Validate

Schema, nulls

Features

Engineer, scale

Train

GBM, 5-fold CV

Evaluate

Holdout metrics

Compare

A vs B

Promote

If better

Interactive dashboard

Five-page Streamlit application for pipeline monitoring and control

Pipeline status

Step-by-step run breakdown with durations
Holdout metrics at a glance
Data validation results

Model registry

All versions with timestamps and metrics
Production flag indicator
Metrics trend over time

Drift monitor

PSI per feature with threshold lines
Distribution overlays (train vs incoming)
Retrain trigger indicators

Model comparison

Champion vs challenger radar chart
Side-by-side metric tables
Promotion recommendation

Retrain

One-click data regeneration and pipeline trigger
Full run history with AUC, F1, and drift flags
Automated promotion decisions logged for audit

Launch live app View source

$ pip install -r requirements.txt
$ python data/generate_data.py
$ python -m src.pipeline
$ streamlit run app.py

Pipeline architecture

Six-stage pipeline with conditional promotion and drift monitoring

Pipeline stages

PSI

Drift detection method

0.85+

ROC AUC on holdout

CI/CD

GitHub Actions

Monitoring

The pipeline monitors data quality and model performance continuously. PSI (Population Stability Index) is computed for every numeric feature by comparing the incoming data distribution against the training baseline. Features with PSI above 0.25 trigger a retraining recommendation. The champion/challenger pattern ensures new models are only promoted when they beat the current production model by a configurable margin on ROC AUC. Every run is logged with full metrics for audit.

Ingest data

Validate schema and nulls

Compute PSI

Per-feature drift scores

Decision

Retrain or keep