Notes
Short write-ups from systems I have built and run: what broke, what it cost, and the number that changed. Each one links to the case study behind it.
How I cut the AI cost of an Arabic OCR pipeline about tenfold
The cost layers of a pipeline that turned 106,595 Arabic posts and 133,161 photos into records: free checks first, OCR once, model tiers, batch pricing.
B2B customer onboarding in the UAE: from 2 weeks to 5 business days
The UAE B2B onboarding sequence I ran, from trade licence to direct debit, and how one owner per step cut it from 2 weeks to 5 business days.
Your construction site cameras already record everything. Can you search it?
Most building sites already record every day on fixed cameras. What plain-language search over the footage takes, and the access checks a buyer should run.
How to check if your looted car in Sudan has been found
Arabati Wein (mycarsd.online) is a free registry of cars looted in the Sudan war. How to search it by chassis or plate number, and what to do with a match.
Why Arabic looks broken in your Facebook data export, and the one-line fix
Arabic in a Facebook or Instagram download shows as ال because Meta writes each UTF-8 byte as a \u00XX escape. Tested Python and Node code to fix it.
Does pay-at-table work in Kenya? What an outside-in diligence found
M-PESA till fees, free amount-filled QR codes and Pesapal Pay@Table: an evidence-tagged read on whether a pay-at-table app can work in Nairobi.
How we cut device-subscription default from 27% to 8%
How I cut default on a UAE device-subscription book from 27% to 8%: a weighted scoring model, AECB checks, signed thresholds and a quarterly refresh.
Why your automation fails silently, and the checks that catch it
Every real outage in my automation exited 0. Five checks catch it: varying counts, canary queries, per-lane quiet limits, transition alerts, repair proof.
A two-key gate for AI agents: one model proposes, a second signs
How I gate AI agent actions: a Claude model proposes and signs, OpenAI Codex co-signs blind, disagreements wait for me, and a ledger keeps every step.
How to catch silent data loss in an AI extraction pipeline
A per-post completion metric passed while about 130,000 records were lost. The four gates that catch it: coverage, grounding, no-shrink and shortfall.