Can recurring LLM traces be synthesized into deterministic pipelines of typed ML and NLP operators? [D]

We are investigating whether recurring LLM workloads can be replaced, where appropriate, by automatically constructed pipelines of regexes, deterministic parsers, traditional ML and NLP models. As an example, suppose an application repeatedly asks a frontier model to read an annual report and return