Built it myself. The filing structure is somewhat consistent enough to parse programmatically, then an LLM handles the unstructured parts (with a lot of trial and error)