MLOps Deployment of a scoring model in production

Python, Machine Learning, MLOps, MLflow, LightGBM, XGBoost, Scikit-learn, Credit Scoring, Optuna, Feature Engineering

Deployment and Monitoring of a Scoring Model (MLOps)

Overview

This project implements the entire production cycle of a Machine Learning model, from its exposure via an API to its monitoring in production. The objective is to transform a previously developed scoring model into a robust, testable, containerized, and automatically deployable service.

The project applies MLOps best practices to ensure the quality, reproducibility, and maintainability of a Machine Learning system in production.

Objectives

• Deploy a scoring model as an API.
• Containerize the application with Docker.
• Automate testing and deployment via a CI/CD pipeline.
• Implement a monitoring and Data Drift detection system.
• Optimize inference performance.
• Document the entire solution.

Demonstrated Skills

ML API Development

• Creation of a REST API with FastAPI (or Gradio).
• Single model loading at startup.
• Input data validation.
• Error handling.
• Automatic documentation via Swagger/OpenAPI.

Automated Tests

• Unit tests with Pytest.
• Error case validation:
  • missing values
  • incorrect types
  • out-of-bounds values
• API stability verification.

Containerization

• Creation of a Docker image.
• Reproducible environment.
• Simplified deployment across different environments.

CI/CD

Full automation with GitHub Actions:

• dependency installation
• test execution
• Docker build
• automatic deployment
• secret management

ML Monitoring

Collection of production metrics:

• response time
• inference time
• score distribution
• error rate
• structured logs
• model inputs / outputs

Data Drift Detection

Automatic analysis of production data to detect:

• input variable drift
• prediction evolution
• statistical anomalies
• potential model degradation

Tools used:

• Evidently AI
• Streamlit dashboards

Performance Optimization

Performance analysis through profiling:

• bottleneck identification
• inference time optimization
• latency improvement
• performance impact validation

Architecture

               GitHub
                  │
                  ▼
          GitHub Actions
                  │
      ┌───────────┴───────────┐
      │                       │
      ▼                       ▼
   Pytest                Build Docker
      │                       │
      └───────────┬───────────┘
                  ▼
             Deployment
                  │
                  ▼
             FastAPI API
                  │
          Model Loading
                  │
        ┌─────────┴─────────┐
        ▼                   ▼
  Predictions          JSON Logging
                              │
                              ▼
                     Log Storage
                              │
                              ▼
               Monitoring & Data Drift

Technical Stack

Machine Learning

• Scikit-Learn
• MLflow

API

• FastAPI
• Swagger / OpenAPI

Tests

• Pytest

Containerization

• Docker

CI/CD

• GitHub Actions

Monitoring

• Evidently AI
• Streamlit

Versioning

• Git
• GitHub

Applied MLOps Best Practices

• Code versioning
• Clean Git history
• Automated tests
• Continuous deployment
• Containerization
• Documentation
• Input validation
• Production monitoring
• Data Drift detection
• Performance optimization
• Environment reproducibility

Results

Upon completion of this project, the model is capable of:

• providing predictions via a REST API;
• being automatically deployed thanks to a CI/CD pipeline;
• being executed in a Docker container;
• being supervised through monitoring metrics;
• automatically detecting data drift;
• being optimized and profiled to reduce inference times.

Acquired MLOps Skills

• Deployment of Machine Learning models
• Development of prediction APIs
• Docker
• GitHub Actions
• Automated tests
• ML model monitoring
• Structured logging
• Data Drift detection
• Performance optimization
• Industrialization of a Machine Learning project