The challenge
The head of settlement operations needed to move break identification from end-of-day reconciliation to intraday — a change that required detecting developing problems before they became end-of-day incidents. Settlement breaks were currently identified only at close, by which point they had already affected the funding line for the next operating day. Each break was fully traceable in hindsight, but the operations team had no capability to see a developing pattern inside the operating day.
Discovery
The operations team's own incident log was the most valuable dataset in the engagement. Every break over the previous 18 months had a forerunner: a corridor latency drift, a counterparty pattern change, a switch-level retry spike. The precursor pattern was visible if you knew what to look for — and the historical record was rich enough to train a model that could look for it systematically.
Our approach
Detection architecture: built a streaming anomaly detector over the settlement event bus with separate detection heads per corridor and per counterparty class
Duty-officer intelligence layer: connected each detection to a context package the operations duty officer can act on within three minutes — the corridor, the counterparty, the suspected cause, the funding-line exposure
Governance design: added a confidence-and-explanation layer so the duty officer can escalate, monitor, or dismiss with a clear audit trail
Runbook integration: connected the detection output to the existing operations runbook so the response is structured and consistent rather than ad hoc
The Outcome
The engagement delivered a real-time anomaly layer on the settlement event bus, with per-corridor detection heads and a duty-officer console surfacing each detection with context, cause hypothesis, and funding-line exposure. The operations runbook governs the response; the model identifies the developing problem before it becomes an end-of-day incident.
The results
“We moved the conversation about breaks from the morning reconciliation meeting to the operations console. That is the meeting we wanted to stop having.”
