Projects
A short, honest list. Two projects, both finished, both documented end to end — including the parts that didn't work.
Leakage-clean ML pipeline on 2,008 patients: five models compared, 143 final predictors, SHAP-driven explainability.
Primary outcome: 6-month mortality at 0.710 AUROC on the held-out Zhang cohort and 0.599 on MIMIC-IV without retraining.
How people arrange cards on an 8×8 board — across 845 recorded trials, with an interactive viewer for retry tracing.
Organized layouts correlated with completion; strategy quality added explanatory power beyond time-on-task. Live explorer for retry tracing.
Because a portfolio of ten unfinished experiments says more about enthusiasm than judgement. Each project here shipped with a documented method, a checked pipeline, and a written account of its limits.
More is coming: clinical survival modelling and behavioural tooling from the MSc, plus the day-job analytics work that pays the rent. I'd rather publish four honest pieces than forty polished claims.