Cut investigation effort about 60% on the Amazon Pay latency project
While the org moved Amazon Pay top-up fraud checks to a new engine to cut latency, I reworked the tool that compares old and new fraud variables so engineers on two teams and the ML scientists could find and fix mismatches fast.
~60%
less engineer effort
~86% / ~89%
fraud-check latency cut at p90 / p99
team launch, June 2018
- Without it
- Engineers hunt by hand for why the new engine disagreed with the old one, and the migration slows.
- The call
- Make the tool generic and backward compatible, with filters and sample orders attached to every mismatch.
- What I did
- Added customer-group and variable filters, sample order IDs for each mismatch and precision control, kept existing jobs running, and sent daily reports to the ML team.
- What changed
- Investigation effort dropped about 60%, my manager said the changes helped both teams, and the team's launch later cut fraud-check latency about 86% at p90.
What people said
“Thanks for improving the tool. I think the changes you have made are generic to help both teams”
“Big customer win!”