Problem
A freelance gift-card client was losing money to fraudulent transactions and needed a more systematic way to catch it than manual review could provide.
Cutting Fraud on a Gift-Card Platform With a Trained Detection Model
Case studyContext & Data
We had access to user information, transaction records, and a set of transactions the client had already identified as fraudulent — a labeled dataset that made a trained model a realistic option rather than a purely rules-based approach.
Approach
The team identified a suitable model, trained it against the labeled fraud cases, and deployed it into production. My own role centered on data prep — preparing and querying the underlying transaction and user data in SQL so the model had clean, relevant signal to train and run on.
Tools
SQL for the data prep work I focused on, and Python for the team's model building, training, and deployment.
Analysis & Execution
The most important work wasn't the modeling itself — it was understanding who the platform's actual average user was, which turned out to be different from our initial assumption. Getting that picture right, largely through the data prep work, shaped which signals the model could actually rely on.
Result
The model reduced fraud during its test period by tens of percent.
Learnings
Understanding the client's needs and who the real end user actually is matters as much as the modeling technique — the average user is often not who you initially assume, and that gap is where a model either works or doesn't.