كتاباتي عن الهندسة المعزَّزة بالذكاء الاصطناعي، وجودة الوكلاء، والقيادة التقنية. الملخّصات هنا، والمقالات كاملةً على Medium.
المقالات منشورة باللغة الإنجليزية.
A purchase step that took seven and a half seconds contained 521 milliseconds of actual work. The rest was queue wait. Five case files on the infrastructure half of a regulated trade pipeline: splitting queues, sizing a worker fleet, and the deploys that quietly skipped it.
Before the story of scaling from ten orders to five thousand, here is the machine itself: a commodity Murabaha as a state machine, where every transition is legally load-bearing and stock has to exist as individually coded units.
The first batch test we ever ran froze three of 31 orders for over twenty minutes; the ten-order retest still lost four. Eighteen months later the same platform ran 5,200 orders end to end at 99.94%. Ten case files on what MySQL did under load, and what it took to keep the pipeline compliant and moving.
It's 2 AM and Slack is exploding: your AI agent is answering payment questions wrong. You fix the prompt, run a few manual tests, deploy — and by 6 AM three unrelated features have broken. Here is how LLM judges catch that before your users do.
