Our task is to develop a machine learning pipeline for a financial institution specialised in consumer and personal loans. This pipeline consists in:
- Implementing a machine learning model able to predict how capable each applicant is of repaying a loan.
- Making the results of the model understandable and available to a non technical audience (namely the customer service, which delivers the news of acceptance or rejection of the loan to the applicant).
In order to achieve this, we trained a number of ML protocols, compared their performance and chose the best. Next, we integrated the model in a dashboard interface deployed as a web application, available at this link. This dashboard resumes the most relevant features of each application, returns the assessment corresponding to the selected ID in real time and uses SHAP values to highlight the features having the highest impact in the assessment.
The frontend is in Python, contains the Streamlit package, and is deployed on Streamlit Cloud, while the backend is written in FastAPI and Python and is deployed on Heroku.
For more details, please check the Presentation folder (presentation and methodological essay). Although this material is currently written in French only, the most essential parts may be understandable even to non-french speakers, for example the table of ML models' performances in the methodological essay.