We use OpenAI/GPT for NLP only - parsing intent - oversimplifying but normalizing things like "I want a serverless function that connects to RDS" is translated to a JSON schema that says serverless <-> RDS. That's then passed to the InfraCopilot engine which then unpacks what that means into a larger architectural representation (we have an Architecture-as-Code blog post coming soon).
This is open source to a large degree, it's powered by the Klotho engine ( https://github.com/KlothoPlatform/klotho )