I'm one of the authors of this blog post. Figuring out how to integrate the LLM, the inference and serving engine, and the constrained decoding implementation such that you get the intended output and overall quality definitely requires careful engineering and testing. Please feel free to AMA. :)