I built this. The interesting part was to force the agentic solution to actually converge because we baked so many oracles and rubrics as gates (validators) into the documentation factory itself that some times the AI was not actually able to find the solution. That basically forced me to create a 6 level intelligence baked into the orchestration itself, yes you read well, 6 different models to audit the output once we hit a validation problem.