01
Prepare your draft
The prototype detects HKID numbers, phone numbers and email addresses in text on your device. You can select other details, such as a person’s name, to mask them yourself.
LEGAL AI, WITH A REVIEW STEP
See what leaves.
Before it leaves.
Privi helps you replace sensitive details with pseudonyms, inspect the text an AI model will receive, and decide when to send it. Your answer can show the real details again on your device.
See how it works
Desktop product in development
A look inside the review step
Review before send
YOUR DRAFT Stays visible to you
Draft a letter for Wong Tai Sin about the unpaid invoice. His HKID is A123456(3). Reply to wong@example.com.
MODEL RECEIVES After approval
Draft a letter for [Person A] about the unpaid invoice. His HKID is [HKID-1]. Reply to [Email-1].
Illustrative example. In the current prototype, you select names to mask; structured identifiers are detected automatically.
The important moment is the one before sending.
THE FLOW
Three steps. You control the middle one.
Privi puts a visible checkpoint between your draft and the model.
02
Review the exact text
Compare your source with the masked version. Add a mask when something was missed. Everything not highlighted is sent as written, so this review matters.
03
Approve, then read
Only after approval does the demo send the masked text to an AI model. Pseudonyms in the reply are restored locally, and you can switch back to see the model’s words.
CLEAR BY DESIGN
Know the boundary.
Masking makes AI use more deliberate. It does not turn every document into safe text automatically.
01
What the model gets
The text you approve, with highlighted details replaced by pseudonyms. Unhighlighted facts, dates, amounts and legal language go through as written.
02
What the prototype does today
Masking and restoration run in the browser. The current demo sends approved masked text through a local development proxy to an AI provider. It is a prototype, not yet a product for live client matters.
03
What requires your judgment
Automatic detection can miss sensitive details. Review every outgoing word before approving it. This is not a substitute for legal review or formal redaction.
WHERE WE ARE
Built for small firms.
Still being built.
Privi is being developed for boutique law firms in Hong Kong. The working browser prototype tests the central idea: show the user what the model will see before anything is sent.
The desktop app, document review, saved matters and automatic detection of names and companies are next-stage work. We’ll update this site as those capabilities become real.