Training and Model Selection
Multiple strategies were compared to prioritize fraud recall while still tracking AUC, F1, and class imbalance behavior.
Key Result
0.9998Fraud recall on the selected run
XGBoostBest-performing selected model
KubeflowConditional workflow and retraining pipeline
GrafanaMonitoring and alert-driven response path
System Evidence
Kubernetes Deployment
The inference API was structured as a deployable service with namespace, resource, and service configuration evidence.
Monitoring Stack
Prometheus and Grafana were used to track system health, metrics, and model-facing behavior after deployment.
Alert-Triggered Retraining
Monitoring signals were wired into an operational retraining story through Alertmanager and Jenkins integration.
Why This Repo Matters
- It demonstrates more than modeling by showing full ML lifecycle thinking
- It ties together data preparation, evaluation, deployment, monitoring, and retraining
- It presents evidence of a production-style workflow instead of a notebook-only experiment
- It is one of the strongest MLOps-oriented projects in the portfolio