How can I safely process sensitive client documents (like investor/loan documents) before sending data into an LLM?

One approach discussed was building a local Python script that acts as a cleaning/pre-processing step: it extracts only the specific data points needed (e.g., interest rate, amortization term, balloon payment) and strips PII before anything goes to the LLM. As long as that script runs locally and is truly air-gapped from the LLM, it doesn't need to touch the model at all — you could even use an LLM to help build the script initially, then run it locally afterward.