The challenge
SummitEdge is a mid-size banking group with reporting obligations to multiple regulators. Every month, a team of eleven analysts pulled extracts from six source systems, reconciled them in spreadsheets, and assembled returns by hand. The process consumed around 400 hours a month and still produced errors.
The risk was bigger than the cost. Two late submissions in one year had drawn regulator attention, and every return depended on institutional knowledge held by a handful of people. When the most senior analyst resigned, the head of compliance called us the same week.
The approach
Regulatory reporting is a data problem wearing a compliance costume. So we treated it like one. Before writing any code, we catalogued every return, traced each figure back to its source system, and wrote down every manual adjustment the analysts made and why. Those adjustments, it turned out, were where both the hours and the errors lived.
- Documented all 23 recurring returns and traced every line item to its source
- Built a single validated data store as the one place returns are computed from
- Encoded the manual adjustments as explicit, versioned transformation rules
- Used machine learning only where rules could not reach: classifying free-text transaction narratives
- Kept a human sign-off step on every submission, by design
What we built
The system is a scheduled pipeline that ingests from the six source systems nightly, validates and reconciles the data, and assembles each return automatically. An NLP model classifies transaction narratives that used to be coded by hand, with a confidence threshold: anything uncertain goes to a review queue instead of into a return.
- Nightly ingestion and reconciliation across six source systems, orchestrated with Airflow
- A validation layer with over 300 automated checks that flag anomalies before they reach a return
- An NLP classifier for transaction narratives, with human review below the confidence threshold
- Auto-generated returns in the exact formats each regulator requires
- A full audit trail: every figure in every return traces back to source records and rule versions
- A sign-off workflow so compliance approves and submits, with nothing sent automatically
400 hrs
of manual work per month, before
38 hrs
per month now, mostly review and sign-off
96.4%
classifier accuracy on transaction narratives
The outcome
Since go-live, SummitEdge has submitted every return on time. Manual effort dropped from roughly 400 hours a month to under 40, a 90% reduction, and those remaining hours are review work rather than data assembly. The team of eleven now runs at six, with the rest redeployed to risk analysis roles the bank had struggled to fill.
The audit trail has proven as valuable as the automation. When the regulator queried a figure in a quarterly return, the team produced the full lineage, from source records through every transformation, in twenty minutes. Under the old process that answer would have taken days, if it existed at all.
