Mohammad Erfan Rashidi

This project tackles a classic customer conversion challenge: predicting which bank clients will subscribe to a term deposit after a direct marketing campaign. Using historical call data from a Portuguese bank, I engineered a set of binary indicators that capture the most predictive signals—such as previous campaign successes, favorable contact months, and specific job categories—while maintaining interpretability. After exploring the data and validating against a dummy baseline, a Random Forest classifier was chosen for its strong balance of accuracy (ROC AUC 0.906) and recall on the minority class, ensuring that high-potential leads are not overlooked. The trained model is packaged into an interactive Streamlit application, making the prediction logic accessible to non‑technical stakeholders. Users can either input client details manually for an instant subscription forecast, or upload a CSV file for batch processing. The app automatically applies the same cleaning and feature‑engineering steps, scales numeric inputs, and returns predictions along with downloadable results. This end‑to‑end pipeline—from raw marketing data to a business‑ready web interface—demonstrates the full cycle of a machine learning solution designed for real‑world decision support.

This project predicts national life expectancy using health and socioeconomic indicators from the WHO Global Health Observatory. After rigorous data cleaning and exploratory analysis, I engineered interaction and polynomial features—such as mortality‑BMI interactions and log‑transformed HIV rates—to capture non‑linear relationships. Multiple regression models were tested, and a Random Forest achieved an R² of 0.958 with an RMSE of 1.93 years, explaining over 95% of the variance in life expectancy. The trained model is deployed as a FastAPI web application, exposing a REST endpoint that accepts JSON input and returns an instant prediction. A lightweight HTML frontend allows non‑technical users to adjust sliders and see the expected life expectancy update in real time. This end‑to‑end pipeline—from raw WHO data to an interactive API—demonstrates the full cycle of building, validating, and serving a high‑accuracy regression model for public health insight.

This project segments credit card customers into five distinct behavioural clusters using K‑Means clustering. After rigorous data cleaning, I engineered domain‑driven features—such as cash advance ratio and purchase frequency—and refined them through correlation analysis, skewness treatment, and multimodality checks, retaining four highly interpretable dimensions. K‑Means with 5 clusters was chosen after comparing silhouette scores and Davies‑Bouldin indices with DBSCAN, and the resulting segments reveal clear profiles, from 'Everyday Super Users' to 'Cash‑Advance Dependent', each accompanied by actionable business strategies. The trained model, scaler, and segment profiles are serialized and served through an interactive Streamlit web application. Users input a new customer’s transaction behaviour, and the app instantly classifies them into a segment, displaying a descriptive profile and a tailored strategic recommendation. This end‑to‑end pipeline—from raw transaction data to a business‑ready decision tool—demonstrates how unsupervised learning can drive targeted marketing, risk management, and product personalisation without friction.