> log parsing

There’s your problem right there. You’re lighting money on fire using an LLM to do a token-intensive deterministic task.