Most engineering teams that put a large language model into a live loop end up building the same scaffolding: a prompt that begs for JSON, a parser with a fallback branch, a retry wrapper, a schema validator, and a small pile of regexes that translate model output into something the rest of the system can