PrivateRedact: offline PII redaction with local LLM, no cloud upload

PrivateRedact addresses a fundamental tension in document privacy: conventional cloud redaction tools solve the technical problem of PII detection by introducing a security problem—the need to upload the sensitive document to a third party. Conversely, existing offline tools either force slow, error-prone manual redaction or apply crude whole-line blackouts that leave the underlying text layer intact and recoverable. The application bridges this gap by combining pattern-based detection (regex for known formats like SSNs, credit card numbers, IBANs, and dates of birth) with semantic detection via a local LLM (for diagnoses, names, addresses, and codenames), then exporting a rasterized PDF where redactions are baked into the image layer itself—making un-redaction or text recovery technically infeasible.

The tool's design choices reflect security-first principles: metadata is stripped by default (PDF Info/XMP, DOCX properties, image EXIF), redacted PDFs can optionally be encrypted, and the app includes a post-export leak scan that re-OCRs the output to confirm no PII survived. The licensing model—free trial for all detection and review work, paid export only—lowers the barrier to testing against real documents, while the one-time purchase (via Gumroad) avoids subscription friction for privacy-conscious users. System requirements are modest (~8 GB RAM minimum), and the app auto-selects the local model size based on available memory, removing a technical barrier for non-expert users.