Cash-flow forecast copilot
the engagement
client data — not shown
heard
"Group treasury asks me for a thirteen-week view and four analysts spend a week rebuilding it by hand."
real problem
The forecast was assembled manually from ledger extracts, AR ageing and a set of assumptions living in one analyst's workbook. It was never reproducible and never explained.
system built
A ledger and AR/AP ingestion layer, a payment-timing model per customer segment, and a retrieval-backed assistant that answers "why did this line move" against the forecast run and its inputs. Every answer cites the ledger rows.
what broke
The assistant answered questions it had no basis for. It now refuses outside the indexed run and says so in plain words.
the open rebuild
same architecture · public data
dataset
Online Retail II + synthetic AR ageing generated from it
what the rebuild covers
Payment-timing model and the cited-answer assistant, with the refusal behaviour tested.
artifacts
NotebookCodeDatasetWrite-upDemo