Projet_07_OCR_Data_Scientist

Credit Scoring Model Deployment

Overview

This project focuses on building and deploying a credit scoring system to estimate whether a client is likely to repay a loan.

It covers the full workflow of a beginner-friendly machine learning product:

• exploratory analysis
• model comparison
• experiment tracking with MLflow
• API deployment with FastAPI
• a simple Streamlit interface
• and local interpretability with SHAP

---

Dataset

The project uses credit application data built from:

• the main application table
• and additional customer-related data sources merged during preprocessing

For deployment, the final API workflow is based on:

• application_train.csv
• application_test.csv

The target variable is the loan repayment / default outcome.

---

Method

The workflow of the project was:

1. explore and prepare the credit data
2. build reusable preprocessing functions
3. track experiments with MLflow
4. compare several classification models
5. evaluate them with AUC, confusion matrix, and a custom business score
6. select a final model
7. deploy the scoring pipeline through MLflow serving + FastAPI + Streamlit
8. add SHAP explanations for local interpretability

---

Key Results

The project led to four main outcomes:

1. Several models were compared, including Logistic Regression, Random Forest, LightGBM, and XGBoost.
2. The evaluation was not limited to standard metrics: a custom business score was defined to penalize costly mistakes differently.
3. The final deployed pipeline is based on LightGBM, which is also used in the API for SHAP-based explanations.
4. The project was turned into a small end-to-end application with:
   • a model served through MLflow,
   • a FastAPI prediction service,
   • and a Streamlit GUI for CSV upload and prediction display.

Overall, this project shows how a credit scoring model can move from experimentation to a simple deployed product.

---

Tools

• Python
• Pandas
• NumPy
• Scikit-learn
• LightGBM
• XGBoost
• imbalanced-learn
• MLflow
• FastAPI
• Streamlit
• SHAP
• Pytest
• Docker
• GitHub Actions
• Jupyter Notebook

---

Repository Structure

.
├── .github/workflows/
├── api/
│   ├── Dockerfile
│   ├── main.py
│   └── requirements.txt
├── data/source/
├── gui/
│   ├── Dockerfile
│   ├── main.py
│   └── requirements.txt
├── mlflow/
├── modules/
├── notebooks/
│   ├── 0_EDA_MLFlow.ipynb
│   ├── 1_modelisations.ipynb
│   ├── 2_modelisations-cuML.ipynb
│   └── 3_deployment.ipynb
├── tests/
├── Présentation.pdf
├── setup.py
└── README.md

---

How to Run

Clone the repository:

git clone https://github.com/jeromelegal/Projet_07_OCR_Data_Scientist.git
cd Projet_07_OCR_Data_Scientist

Then install the required dependencies for the API or the interface depending on the component you want to run.

Example:

pip install -r api/requirements.txt
pip install -r gui/requirements.txt

You can then:

• explore the notebooks with Jupyter
• run the API with FastAPI / Uvicorn
• and launch the GUI with Streamlit

---

Main Takeaways

This project helped me practice:

• classification on imbalanced data,
• business-oriented model evaluation,
• experiment tracking with MLflow,
• model serving,
• API development,
• simple front-end integration for machine learning,
• and prediction interpretability with SHAP.

It was also a strong introduction to the deployment side of a data science project.

---

Limitations

This project has a few important limitations:

• the interface remains simple,
• deployment is still closer to a learning / portfolio setup than full production,
• and model performance depends heavily on preprocessing and class imbalance handling.

A natural next step would be to improve monitoring, threshold tuning, and production robustness.

---

Author

Jérôme Le Gal
Data Science student