How can I build a tool to strip personal information out of client PDFs before using them with AI?

One member built a local Python program (developed with Claude's help) that redacts PII from PDFs before they're used with AI tools. Unlike using Claude directly to redact text (which just regurgitates it into a separate text file), the Python script resaves the actual PDF with black redaction marks over sensitive fields. It runs as a local installation on the computer rather than online. The tool can automatically detect and redact things like bank routing numbers, social security numbers, addresses, and zip codes, but it does not automatically know client names since there's no standard format for where names appear in a document—those have to be manually fed in. It could potentially be trained to look in specific locations if you specify a document type like a 1040 or 1120S. Another member noted a similar approach discussed elsewhere: have Claude scrutinize the Python script multiple times to check its logic, since LLM outputs can vary between runs, and repeat the check until you consistently get the same clean result. The tool is described as self-contained and airgapped, meaning it doesn't send data out to the main AI model.