Senior Data Scientist with 6+ years across the New Zealand public sector (health and analytics) - working end-to-end from messy operational data through to production analytics, machine learning, and decision-ready reporting. Māori (Ngāti Apa). Strong in R, SQL and Power BI, and in Python end-to-end across the projects below, with hands-on work in applied ML/NLP, retrieval-augmented generation, and locally-run large language models in regulated, privacy-sensitive settings.
They are public-data portfolio projects that demonstrate techniques I use in professional work. Work projects use confidential data and can't be published, so these examples are built on open datasets - the methods, the rigour and the decisions are the same.
-
clinical-guideline-rag: retrieval-augmented question answering over UK NICE clinical guidelines, running entirely locally (Ollama + Chroma): grounded answers that cite the exact recommendation and refuse when the guidelines don't cover the question, with a RAG evaluation harness for retrieval quality and faithfulness.
-
hospital-readmission-risk: an end-to-end, equity-aware 30-day hospital readmission model: leakage-aware data preparation, probability calibration, SHAP explainability, a fairness audit across demographic groups, and a FastAPI service with tests and CI.
-
smoking-cessation-effect: causal inference: the effect of quitting smoking on weight gain, estimated with g-computation, inverse-probability weighting and a doubly-robust estimator, plus overlap/balance diagnostics and an E-value sensitivity analysis.
-
flu-forecasting: short-term influenza forecasting evaluated with rolling-origin backtesting against honest baselines; a gradient-boosted model beats seasonal-naive, with a regime-aware read on the COVID-era disruption.
-
clinical-note-classifier: classic ML vs a local Llama on clinical-text classification: a numbers-backed look at when a trained model beats an LLM, and when it doesn't.
-
health-analytics-dbt: analytics engineering: a tested, documented dbt + DuckDB pipeline turning raw health records into layered staging and mart models with data-quality tests.
-
health-dashboard-powerbi: an interactive Power BI dashboard on top of those dbt marts: a star-schema model, DAX measures, and coordinated multi-page report, shipped as a text-based project (
.pbip) alongside the built.pbix.
R (tidymodels, Shiny) · SQL & database design · Power BI (DAX, RLS) · Python (pandas, scikit-learn) · machine learning · calibration & SHAP explainability · causal inference · time-series forecasting · NLP & text classification · GenAI / LLMs (local Llama, RAG, LangChain, vector databases, evaluation) · dbt / analytics engineering · equity-aware analytics