GenAI systems, end-to-end ML platforms, commercial analytics and applied machine learning, each backed by source code, a live demo or a written case study.
Case studyA multi-LLM system that collects 2,000+ live roles a day from 9 ATS APIs and job boards, scores each one against my CV with a three-model LLM-as-a-judge cascade, and assists with applications through a Chrome extension that learns from my edits.
End to endAn end-to-end platform on PySpark and Delta Lake: bronze, silver and gold layers, star-schema marts, and churn, fraud and high-cost-claim classifiers with MLflow experiment tracking, batch scoring, data-quality monitoring, Docker and CI/CD in GitHub Actions.
AnalyticsA public version of the collections segmentation approach I delivered at Infosys, on synthetic data: RFM features engineered in SQL with CTEs and window functions, high, medium and low risk tiers, a recovery playbook and an impact simulation, reported in Streamlit.
MSc dissertationMy MSc dissertation as a reproducible pipeline: 10,000 Amazon Books reviews from the Amazon Reviews 2023 dataset, Bag-of-Words and TF-IDF features, tuned SVM, Naive Bayes and Random Forest classifiers, and an analysis of reviews whose star rating contradicts their text.
MScQuestion answering over business reports, with every answer traced to its source page. Reports in PDF, Word, Markdown or text are chunked, embedded into a FAISS index and retrieved for a Gemini model, with automated answer evaluation and unit tests across the pipeline.
MScThe pricing and promotion methods I used in industry, rebuilt on synthetic scanner data with known elasticities: log-log demand models for 70 market and product groups, baseline and incremental decomposition, a guardrail-constrained promotion optimiser and an elastic-net marketing mix model.
MScMSc Big Data module coursework: a PySpark workflow that engineers features and trains logistic regression and random forest classifiers for customer churn, with the results presented in Tableau.
LiveDeployed
LiveDeep learningA TensorFlow convolutional neural network trained on 2,152 PlantVillage images to classify potato leaves as early blight, late blight or healthy, served through FastAPI, a Streamlit app and a command-line tool.
LiveClassification
LiveRegression
LiveClassification
RegressionA modular training and inference pipeline that estimates gemstone prices from carat, cut, colour, clarity and dimensions. Five tuned ensemble models, including XGBoost and CatBoost, are compared and the best on test R² is served through a Flask API and a Streamlit front end.
RegressionA modular regression pipeline that estimates a student's maths score from demographic and academic indicators. Nine regressors are benchmarked, the best is persisted, and scores are served through a Flask app.