Budget vs Actual Dashboard
The variance numbers are easy. The commentary is the slow part — so let the model write the first draft from the actual figures.
The problem
FP&A teams generate variance reports every month, then write the commentary by hand. The commentary is the slowest, most repetitive part of the cycle — and it reads the same way every month because it is built from the same template.
What it does
An interactive dashboard over 24 months × 3 business units of synthetic data. Filter to a business unit and period, click Generate commentary, and an LLM produces a board-ready narrative — then download it as a PDF.
- 24 months × 3 business units of synthetic data
- LLM grounded on the variance table — no hallucinated numbers
- Downloadable PDF commentary with one click
Design choices
The model is grounded on the actual variance table, not asked to invent numbers. The README explains the prompt that forces every figure in the narrative to trace back to a row in the data. That grounding is the difference between a useful draft and a plausible-sounding fiction.
Three deliberately injected variances in the synthetic GL (a cloud-hosting overrun, a licensing revenue miss, a travel spike) give the commentary something real to find.
Stack
Built and demonstrated on synthetic data only.