One frozen snapshot of federal budget data through three modern stacks: a diffable Power BI model in TMDL, a dbt + DuckDB warehouse with 82 tests — and the five things the tests caught in real data.
A natural-language-to-SQL copilot with no LLM in it: a deterministic router, 40 pre-approved SELECTs, results frozen at build time, and a refusal taxonomy — out-of-scope, untracked-agency, injection, and more, each with a named reason.
Linear continuation vs. damped Holt-Winters, walked forward over 36 held-out periods. The winner’s margin is thin enough that a library convention flips it — so the page shows both methods and says so out loud.
The flow simulator follows a real appropriation — budget authority to obligations to outlays, month by month — and lets you pull the levers. Real data, frozen, with a scenario engine that applies each lever in a fixed, documented order.
A budget workbench in Excel: four template workbooks, six VBA macros for reconciliation, variance, and the close checklist — plus Python twins of the two computational macros so every number reproduces headlessly.