A forecast-vs-actual dashboard that computes variance, accuracy (MAPE), and bias, ranks the biggest customer and product drivers, and writes a grounded plain-language explanation — every sentence maps back to a number on the page.
Forecast-vs-actual reviews are usually a spreadsheet plus a gut-feel narrative; wanted the narrative to be provably grounded in the numbers.
A static analytics dashboard that computes standard FP&A metrics client-side and generates a plain-language summary that only describes what's on the page.
One dependency-free analytics module shared by the UI and the test suite. CSV in, computed KPIs plus narrative and exports out — no server, no account.
Deliberately not wiring up a language model for the "explainer" — every sentence is templated from a real computed value, stated as a contract in both the UI and the docs.
How to design a metrics layer trustworthy enough to hand to a finance audience without a hallucination disclaimer.
Source repository
Back to the portfolio