Production-style, decision-driven projects focused on business impact, ROI-aware modeling, and real-world constraints.
End-to-end NLP pipeline that predicts CVE severity (LOW/MEDIUM/HIGH/CRITICAL) using TF-IDF bigrams + XGBoost and a fine-tuned SecBERT model. Achieves 0.81 F1 on CRITICAL — the highest-stakes class — with SHAP token-level explainability and an LLM cost-benchmark comparison.
Production-aware fraud detection system that reframes detection as a capacity-constrained ranking problem. Uses a calibrated LightGBM + XGBoost ensemble with temporal splits, cost-sensitive thresholding, SHAP explainability, and concept drift monitoring — achieving 81.3% fraud recall at 500 daily reviews with 70% cost reduction.
Profit-optimised retention system that goes beyond churn scoring — combining survival-adjusted CLV, T-learner uplift modeling, and 0/1 knapsack MILP allocation. Achieves 8.15× ROI at £10k budget with correlated Monte Carlo risk simulation across 2,000 scenarios, outperforming highest-churn-first targeting by +37.7% mean profit.
Answers the real operational question for a Portuguese bank's outbound campaigns: given N calls per day, which customers maximise expected profit? Uses Optuna-tuned Random Forest (ROC-AUC 0.81) with SHAP-driven feature analysis, achieving 77% conversion precision in the top 1,000 contacts — 6.8× better than random — with an interactive profit simulation dashboard.
Subscription analytics pipeline that turns churn/usage signals into revenue intelligence for forecasting and retention-focused decisions.
Forecasts total daily sales with proper time-based validation, trend/seasonality diagnostics, and promo impact analysis.
End-to-end pipeline that ingests job postings, normalizes skills, and tracks momentum over time for labor market intelligence.
Mini decisioning + pricing engine that predicts late-payment risk, approves instant pay, and generates a risk-based fee quote.
I'm continuously shipping decision-focused systems and analytics tools. Explore all repositories and recent updates on GitHub.