How local PII scrubbing works
When you type a prompt in Claude, names, emails, phone numbers, and other identifiers can end up on someone else's servers. SafAI intercepts that flow on your device before the prompt is sent.
Detection on your machine
SafAI runs all detection on your device before your prompt is sent. It combines pattern matching (for structured data like emails and phone numbers), local AI-based entity detection, and other on-device analysis to identify persons, organisations, locations, and other identifiers. No cloud API receives your raw prompt for detection.
Tokenisation
Each piece of PII is replaced with a stable token like [PERSON_1] or [EMAIL_1]. The mapping lives in local memory (up to 1000 tokens per session) so follow-up messages stay consistent—if John Smith becomes [PERSON_1] once, he stays mapped across the thread.
What the AI sees
Only the scrubbed text reaches the AI provider. The model responds using tokens. SafAI then restores original values in the response on your screen, so you read natural language while the provider never saw your raw PII.
Why local matters
Cloud redaction services still require you to upload the sensitive text. SafAI's architecture means the sensitive payload never leaves your machine in identifiable form—a different trust boundary than "trust our API to delete it later."