I turn messy records into reports people can act on, with every number reconciled back to SQL before it ships.
ScrollNo report ships until its headline numbers match an independent SQL calculation. Power BI and Tableau are checked against DuckDB, value for value.
A check only counts if it has been seen to fail. Each one is proven by corrupting the data deliberately and watching the build stop.
Sensitive attributes like gender, age and region never become model features, and the fairness table shows what the model does anyway.
Dashboards are generated from code, not clicked together: reviewable, versioned, and rebuilt identically every time.