I'm a data analytics postgraduate in London. I build leakage-clean predictive models, interactive behavioural analysis tools, and reports that stay honest under scrutiny — because a number only helps if it survives the audit.
Featured work
Every piece here is built on data I can defend: documented sources, audited pipelines, and results I can reproduce.
0.71 AUROC · 6-month internal
A leakage-clean pipeline on 2,008 patients comparing five models. XGBoost reached 0.710 AUROC at six months; SHAP explains the cardiorenal axis it learned — and 0.599 transfers to MIMIC-IV without retraining.
845 recorded trials · 714 usable
Given cards — and sometimes blanks — on an 8×8 board, how do people organise what matters? 845 recorded trials show organisation beats raw effort, and blank cards act as scaffolding. Live explorer for retry tracing.
Capabilities
Four moves, repeated carefully: model, clean, explain, report.
Regression, survival models, tree ensembles — benchmarked honestly, with baselines.
Missingness audits, leakage hunting, and feature hygiene before any model sees the data.
SHAP-driven reduction and calibration checks, so stakeholders can see why a model says what it says.
Board-ready dashboards and static reports that separate signal from confidence theatre.
Approach
01
A high AUROC built on leaked features is a liability, not an achievement. I audit first, then report what survives.
02
A model nobody can question is a model nobody can trust. SHAP, calibration and honest failure modes are part of the work.
03
With 2,008 patients or 845 trials, I report what the data can actually support — and say plainly when that's a limit.
· London, UK · Available from 2026 ·