Sounds interesting. Which AI are you using to generate the code? Or is it one you built yourself. Can you confirm the accuracy of all the generated code?
I am self hosting Qwen3 and also use a series of models from OpenAI as a fallback / load balance.
I don't guarantee accuracy of data, however I have built in mechanisms for the AI to critique its own response before returning, which means that it is considered by my users to be accurate enough to be a major part of their production data workflows now.