Fraud Detection Engine

The Spreadsheet Problem

Someone was looking at spreadsheets and flagging suspicious transactions by eye. That’s not a scalable fraud strategy — it’s a slow way to burn out the person doing it. 1.5 million payment transactions, and the system for catching anomalies was mostly manual review.

Rules First, Dashboards Second

The detection logic was straightforward: transaction patterns that deviated from established baselines. Unusual amounts. Frequencies that didn’t match history. Destinations that shouldn’t have been there. Python handled the rules, SQL queried the raw data, QuickSight gave the non-technical teams a view into what was flagged and why.

Where the 80% Came From

The rules caught the obvious cases — transactions that should never have needed a human to review. The remaining 20% was the edge cases: borderline amounts, new patterns, things that needed context a rule can’t encode. That tail is where the real work is. But automating the easy cases meant the team finally had bandwidth to focus on what actually required judgment.

What I Learned

A simple system people trust beats a complex one they don’t understand. Fraud detection isn’t fundamentally a model problem — it’s a signal problem. Clean data, clear rules, and a process for handling false positives. The QuickSight dashboards worked because they told a story, not because they displayed numbers.