My most substantial builds are enterprise-style pipelines spanning data engineering, modelling, optimisation and deployment.
Models how unit sales respond to price, promotion, display, seasonality and competition using log-log regression (price coefficient = elasticity), decomposes every week into baseline vs incremental lift, and feeds the elasticities into a Google OR-Tools optimiser that recommends the most profitable trade-promotion calendar under margin, trade-spend and average-price guardrails. A parallel R / glmnet Marketing Mix Modelling workstream adds contribution decomposition, ROI ranking and scenario simulation.
An enterprise-grade, end-to-end platform that runs 62,000+ insurance records through a Bronze → Silver → Gold medallion architecture raw ingestion, cleansing and validation, then star-schema marts and ML-ready datasets. On the Gold layer it trains and evaluates three decision-support models policy-churn risk, claim-fraud risk and high-cost-claim prediction and produces batch-scored outputs, all driven by config-based runs, automated tests and CI/CD.
Segments a 200,000+ customer lending portfolio by payment behaviour and outstanding debt using an RFM model adapted for collections (recency of last payment, payment frequency and outstanding exposure), engineered in SQL with CTEs and window functions and banded into High / Medium / Low risk. Recovery resources are then aligned to those tiers, and a transparent impact simulation shows the targeted strategy lifting recovery rates ~14% and cutting average recovery time ~12 days versus an untargeted baseline. Ships with an executive Streamlit dashboard and a Power BI / Tableau rebuild guide.
Self-built projects covering the full lifecycle data, modelling, deployment and a live demo where hosted.
An end-to-end regression pipeline (ingestion → transformation → training → inference) that prices gemstones from carat, cut, colour, clarity and dimensions benchmarking scikit-learn, XGBoost and CatBoost and auto-selecting the best model by test R², served via Streamlit and a Flask JSON API.
A reproducible diagnostic classifier on the Wisconsin dataset using a StandardScaler + Logistic Regression pipeline (malignant vs benign), evaluated on accuracy, precision, recall, F1 and ROC AUC (~0.98 accuracy, 0.996 ROC AUC), with a Streamlit app and automated tests.
An end-to-end regression pipeline predicting a student's math score from demographic and academic indicators benchmarking several regressors, persisting the best model, and serving predictions through a Flask web app.
Selected projects from my Master's covering NLP, big-data, GenAI and marketing analytics.
Built and compared text-classification models (TF-IDF features with SVM, Naive Bayes and Random Forest) to classify review sentiment and generate marketing insight.
Built a PySpark + Tableau workflow for churn prediction feature engineering, Logistic Regression and Random Forest with model evaluation on a big-data pipeline.
Designed a GenAI assistant using retrieval-augmented generation to query business reports and return source-grounded, citation-backed summaries.
Python-based simulator estimating sales contribution, ROI and budget-allocation scenarios across pricing, promotion and marketing drivers.