The challenge
Norden's fraud-detection model relied on a constantly-refreshed view of card transactions — millions of rows a day, each carrying PAN-hashed identifiers, merchant metadata, geolocation, and outcome labels. Every external partner who needed to look at the data — model vendors, the FCA reviewer, the consultancy auditing the model — required a fresh data-sharing agreement that took three months to land.
The team's internal vendors also struggled. Junior analysts couldn't get hands-on training data; the production environment was so locked down that no one could prototype outside a remote VDI session.
The solution
Norden deployed Doppelset Lab on a dedicated tenant in eu-north-1, attached to their card-network data warehouse via a read-only replica. A relational twin was trained over four tables — accounts, transactions, merchants, disputes — preserving FK cardinalities and the 0.4% fraud base rate.
Every night, the twin re-trains on the previous day's data and emits 6 hours later a one-week synthetic window. Vendors, analysts, and the regulator query this synthetic window through Snowflake — using the same SQL as production, just pointed at a different schema.
When the team needs balanced data for training a new fraud model, they call twin.sample(rows=100_000_000, balance={'is_fraud': 0.5}) and get a perfectly balanced corpus on demand, with a signed report attached.
Our fraud team finally has a shared dataset everyone — analysts, vendors, the regulator — is allowed to look at. That alone paid for the platform.
The results
Vendor onboarding fell from 92 days to 9
External vendors now receive a synthetic snapshot and the signed receipt under NDA on day one. No more three-month DSA cycles for proof-of-concept work.
A fraud model that holds up
The fraud model retrained on the synthetic dataset, then tested on a held-out week of real data, came within 0.6 AUC of the production baseline. Crucially it retained the rare-fraud-typology long tail.
An audit that didn't bite
Norden handed the FCA reviewer the signed Doppelset receipts alongside the model card. Review wrapped in eight days — Norden's fastest ever.
Internal experiments unlocked
Junior analysts now prototype fraud-detection ideas in regular notebooks against the synthetic warehouse. Number of new fraud signals tested per quarter jumped from 4 to 27.
What's running
Next story
Pulse Telecom — telecom & churn