Two local AI models plus checksum-validated pattern matching scan every message in about 70 milliseconds, before anything leaves your machine. Nothing is sent to the cloud to be "checked."
As you write, an on-device ensemble (two ONNX models plus Luhn- and checksum-validated patterns) flags every identifier: names, SSNs, accounts, addresses. Nothing has left your machine yet.
Each real entity gets a deterministic alias that stays consistent across every conversation. Structured IDs become shape-preserving surrogates, so the AI can reason normally without ever knowing who anyone is.
Confident mode sends directly; Review-first shows the exact outbound payload (byte-for-byte, with its hash) behind an Approve & Send gate. What you approve is literally what is sent. Every send produces a hash-chained receipt.
The model replies about Client_1; Amnesia AI® swaps the real names back in on your device. You read a normal answer about your actual client. If anything is ambiguous, the send is blocked: fail-closed, with a receipt explaining why.
25+ PII types. Money amounts and ordinary dates are preserved: the AI needs them to reason.
Names, organizations, emails, phones, addresses, dates of birth
SSN, EIN, driver's license, professional license numbers
Credit cards (Luhn-checked), bank accounts, ABA routing, IBAN, crypto wallets
Canada SIN; UK NI, NHS, UTR; Australia TFN, ABN, Medicare; IPs, URLs, government IDs
Structured IDs get shape-preserving synthetic surrogates (an SSN becomes a never-issued number) so the AI's output stays coherent. Each real entity keeps the same alias (Client_1, Company_1) across every conversation, for years.
The live Shield Test panel shows detection in action, on your machine, offline.