Reconciliation that closes in minutes instead of days

The challenge

Three systems disagreed about what had been paid, and the difference was found by a finance team exporting CSVs and sorting them by hand. Month-end close took days, breaks were discovered long after the counterparty could be asked about them, and nobody could say with confidence what the real exposure was.

What we did

1. 01Streamed ledger, processor and bank events into one normalised transaction record keyed on a stable identifier
2. 02Made matching idempotent and replayable, so a late or duplicated file corrects the picture instead of corrupting it
3. 03Classified breaks by cause — timing, fees, FX, genuinely missing — so the finance team triaged rather than searched
4. 04Gave every number a drill-through to the source events that produced it
5. 05Alerted on unmatched value crossing a threshold rather than waiting for the close