Been working on local-first LLM observability for exactly this use case — tracing local model pipelines without sending data to cloud. Happy to share if anyone's interested.