Governed AI Software Factory

Challenge

The company wanted one governed route from an idea to working software for its own delivery teams. The route had to produce the documents a customer expects, from brief and requirements to architecture, design, plan and proposal, in English and Arabic, and then hand bounded coding tasks to AI agents. Every approval had to belong to a person, and every claim had to be traceable to evidence.

Our approach

We built the pipeline in ten weeks. People work through eight stages in a web workspace, in English or right-to-left Arabic. A language model drafts each document as structured data; the system validates it against a contract and its own rule checks, and saves it as a proposal. Every requirement must quote the source text it came from. Each stage ends at a gate that freezes the package it hands over, and a gate goes on hold if an approved input later changes. From task release on, coding agents receive one bounded task with the files they may touch and the tests they must pass. Each patch is applied to a clean copy of the repository and tested before a person reviews it. Agents and the model have no way to record a decision. Artifacts are identified by their hash, approvals bind to an exact version, and evidence tables refuse edits and deletes. In every recorded evaluation run, the model caused no decision that a person had not made.