This article is part of the Agents that touch production series. The code and data for the rebuild are in the repository linked at the end.
The problem in plain terms
Before any modelling: here is the business situation and why the obvious approach fails.
What the data actually says
What you can see when you look at the numbers carefully, and what you cannot see — and why the gap matters.
The method
The approach chosen, and why alternatives were ruled out. Assumptions stated, not buried.
What shipped
How this ended up in production, what it took, and what the first version got wrong.
Run it yourself
All notebooks and code in the repository are runnable from a clean environment.
git clone https://github.com/emadhsnbuilds/fraud-review-assistant
cd fraud-review-assistant && pip install -r requirements.txt
jupyter lab