My medical text classifier scored 100% on its test set, so I built an independent stress test, and it scored 18%. I used LIME to find the causes, caught a data leakage bug of my own, and raised real reliability to 92.3% over 12 iterations. I also added a confidence threshold so uncertain cases go…
I loaded 901M NHS GP appointments into Snowflake and modelled them with dbt, with 7 tests, including a check that no records are lost. Appointments booked over 4 weeks ahead are missed 5 times as often as same-day ones (9.31% vs 1.83%), and face-to-face appointments are missed twice as often as…
A self-started project. I built a Python pipeline over 2 years of half-hourly Elexon data (35,663 periods) and checked the duplicates and revisions before trusting any numbers. Prices went below £0 in 1,950 half hours. They peaked at 29% of June 1pm periods and were 3 times as common at weekends…