Engineering, explained
as I ship it
Notes from real production work, payments, AI agents, architecture, and why each decision was made.
Browse by topic
Writing
The alert channel nobody read, and the catalogue that fixed it
Every monitor paged, so none of them meant anything. Staging woke people at night, one quiet tenant could report a 100 percent error rate off a single request, and the team had learned to swipe the notification away. Here is the P1 and P2 split we settled on for an ECS stack, and the two mechanics that stopped the flapping.
Read articleA reader read my benchmark better than I did
I published a fine-tune that beat prompting by 29 points. A commenter said the gap between my two accuracy drops was the real signal. He was right that my table was misleading, wrong about which arm was inflated, and the experiments I ran to check produced a third answer.
The client wanted production data in preprod. I said no, then gave them something better.
A reasonable request from a reasonable client: make preprod look like prod, with real data. Why that would have been the worst security decision in the project, how I pushed back without a fight, and the anonymised seed pipeline that made preprod more useful than a copy would have been.