A churn score has limited value in isolation. The useful work is turning it into a repeatable workflow with inputs, explanations, recommendations, delivery, and checks over time.
The problem
What had to work beyond the model.
Covers ingestion, feature engineering, model training, scoring, and retention recommendations.
Includes explanations and an analytics surface rather than exposing only a raw probability.
Repository documents monitoring, batch entry points, tests, containers, and CI.
How the pieces connect.
- 01Data inputs
Customer, activity, transaction, and support data move through ingestion and feature-engineering stages.
- 02Model candidates
The workflow supports baseline and boosted classifiers, then resolves an active model through a registry pattern.
- 03Decision service
FastAPI exposes scoring, explanation, and retention-recommendation endpoints for downstream use.
- 04Command center
A Streamlit surface and monitoring reports make model behavior, data quality, and retention actions inspectable.
Choices that make the system usable.
- 01
Designed a model registry pattern so the API and dashboard can resolve an active model source instead of relying on a hard-coded artifact.
- 02
Included both a synthetic demo path and a Kaggle Cell2Cell workflow so the system can be inspected without pretending one dataset fits every use case.
- 03
Made per-customer explanations and treatment/control assignment visible rather than exposing an unexplained probability.
- 04
Made monitoring and retention logic visible parts of the architecture, not hidden implementation details.
How the work can be inspected.
- The repository documents batch entry points, tests, containers, and CI alongside interactive analytics.
- Monitoring reports are designed to surface drift, missingness, classification quality, and calibration through Brier score.
- The service separates score, explain, and recommend actions so reviewers can inspect the decision path.
This is an end-to-end engineering system, not a claim of a deployed churn program. No accuracy, retention lift, or A/B-test outcome is asserted without an evaluation report.
PythonFastAPILightGBMXGBoostscikit-learnStreamlitDockerSQLite