What we are writing about.

Slide 2 of 6 · Week one: we just watched.

We sat with the analyst running the reconciliation and wrote down every keystroke. The override sheet had three competing layouts for the same fields. The settlement file changed schema on the first of every month. By the end of the week we had a one-page diagram of every input, every transformation, every output.

Slide 3 of 6 · The architecture we picked.

A small, sensible stack. A scheduled Python job picked up the source files at 5:30 a.m. and wrote them into a Postgres staging schema. dbt models on top of staging produced four reconciled tables. Power BI on top of dbt produced the dashboard the credit committee actually opened.

Slide 4 of 6 · Eleven flagging rules, all in SQL.

The credit committee told us, in three short sessions, what they actually looked for when they read the morning report. Eleven rules. Each one a single SQL statement against the reconciled tables, named and documented. Line utilisation over a threshold. Scheme exposure outside an agreed band. Exception count above a daily ceiling.

Slide 5 of 6 · What sign-off looked like.

By week five, the morning routine for the analyst had become: open the dashboard, scan the exceptions table, ask one or two clarifying questions in Slack, sign off. The reconciliation itself ran without her.

Slide 6 of 6 · The numbers, plus what we did not do.

92% of the manual reconciliation time was removed. Five weeks from sign-off to live. Zero findings from the internal audit walk-through. We did not replace the override sheet, did not introduce a new BI tool, did not write a long policy. We built only what they could maintain themselves.