Selected projects and technical work relevant to fraud detection, churn prediction, credit risk, and ML system design.
Demos and Previews
Portfolio Projects
YesAlps Ski Booking Web App
Problem
YesAlps had no product comparable to modern booking and marketplace experiences.
Approach
I built a Java web app concept inspired by Airbnb/Booking flows, and added a ski-slopes layer by scraping and integrating slope data directly.
Result
An end-to-end prototype combining backend development, data collection, and interactive map navigation.
TIM / CladAG Churn Classification Challenge
Problem
Identify high-risk churn users for TIM, Italy’s largest telecom operator.
Approach
I designed a churn classification pipeline with feature engineering and robust model validation.
Result
I won the challenge.
Kaggle Modeling Portfolio
Problem
Show practical ability to explore datasets and build strong predictive models across different domains.
Approach
I published notebooks covering exploratory analysis, gradient boosting models, and neural network workflows with reproducible evaluation.
Result
Kaggle Code Master, career-high world rank: #163.
LightGBM Inference-Time Feature Importance
Problem
Standard feature importance methods are often disconnected from real-time scoring behavior.
Approach
I implemented and documented a method to compute feature importance at inference time for tree-based models.
Result
Published as open-source code and a technical article.
Production ML Case Highlights (Anonymized)
Insurance Fraud Detection
Scope
Built and deployed fraud-risk scoring to support claim triage.
Delivery
From feature engineering to model training, evaluation, and integration in decision workflows.
Loan Default Prediction
Scope
Developed and deployed default-risk models for underwriting and portfolio monitoring.
Delivery
Risk-score generation with model validation, calibration, and business-oriented thresholding.
Multi-Series Forecasting
Scope
Delivered forecasting systems across multiple time series for planning use cases.
Delivery
Built forecasting pipelines with backtesting and periodic model refresh.