Banks, telecoms, and government bodies across the region want AI's leverage without shipping sensitive data to a foreign cloud. Here's how on-premise, self-hosted AI actually works — and when it's the right call.
While building Labiba AI, a RAG assistant for Zain's Legal Affairs department, the pipeline kept falling apart at step one: reading scanned Arabic letters. Here's how a locally hosted vision model ended up replacing the OCR engine entirely.
Users kept telling me the same thing about Labiba AI: "I know the answer is in this document, but it can't find it." The information was always near the end of a long, multi-page letter. Here's why that happens and how parent-document retrieval fixed it.
The Musaned series finale: owning deployment end to end — automated zero-downtime releases on Linux, Docker images headed for OpenShift, and going from a team of one to technical leadership.
Part 9 of the Musaned series: the technical companion to the one-week launch story — how memos, circulation, voting, and permanent decision records are modeled in Laravel and Filament.
Part 8 of the Musaned series: a business process management tool for upper management — all the way to the CEO — with a two-week deadline, no requirements document, and a launch in one week.
Part 7 of the Musaned series: the flagship feature as a case study — how the business team went from waiting on engineers to searching 2M+ voucher records themselves.
Part 6 of the Musaned series: the architecture that kept countless integrations manageable — one service class per external system, carefully ordered queues, and Redis as the cache store.
Part 5 of the Musaned series: fiber customer information lived in Microsoft SQL Server, and getting PHP to talk to it reliably was its own dependency saga — with different traps than Oracle.
