LightGBM · MLflow · FastAPI · Docker · ONNX

From scoring model to monitored service

A prediction pipeline connecting experimentation, an API, deployment and drift analysis.

The problem

Turn a credit scoring model into a technical solution whose predictions, versions and changes can be monitored.

My role and choices

I worked on experimentation and the MLOps pipeline: prediction serving, containers, continuous integration and data monitoring.

Track experiments and compare models with MLflow.
Expose predictions through a containerised API and automate the pipeline with GitHub Actions.
Analyse drift with Evidently and compare LightGBM and ONNX inference.

Results

The saved benchmark reports a probability difference of 7.31 × 10⁻⁸ between the compared paths. It provides an export comparison, without measuring full HTTP service performance.

Limitations and next steps

The local benchmark compares different computation paths. Cloud latency, concurrent load and current deployment were not checked in the audit. This educational project is not a financial decision service.