Mask personal data before you send the document anywhere
Paste text or drop a file. The page finds identifiers and names, tells you why each one matters, and masks the ones you approve. It all runs in your browser — the document never leaves your machine.
Drop a .txt, .md, .csv, .json or .log file, or click to choose
Scanned locally · no upload · nothing leaves this page
Runs a 66 MB name model in this tab. The first press downloads it once and caches it; after that it works offline. Your document is never uploaded — only the model is downloaded.
Finding is not the same as fixing
Most privacy tools highlight what they find and leave you to edit it by hand — which is where the mistake happens, because a document with forty findings is forty chances to miss one. This one does the whole job: it finds the identifiers, shows each one with the reason it matters, lets you tick what to remove, and then rewrites the document so what you send is a different file.
It then re-scans its own output and tells you whether anything is still there. That check is worth insisting on: a mask that leaves a detectable value behind is worse than no mask, because it looks safe.
What it looks for, and why each matters
Two engines, because they fail in different places:
- NRIC / FIN, email, phone, card, bank account, company UEN, postcode — found by exact patterns. These are structured: a checksum, a fixed length, an
@. No model is needed, and a model would be worse at them. - People, companies and places — found by an on-device name model. These have no pattern. "Kevin Seow is the boss, Chek Wei is the subordinate" names two real people and contains nothing a regular expression can key on. Press Also find names, places and companies to run it.
- Relationships — "the boss of Acme Pte Ltd" identifies two people at once, so the named relationship is caught even when only one name is spelled out.
When both engines run, names that look like the same person — Kevin Seow and Mr Seow — get the
same placeholder. Giving them different ones would quietly tell a reader that two people were involved, which is
the opposite of what masking is for.
What the name model does to your privacy
The model is a 66 MB file that has to be downloaded the first time you press the button. That request goes to a model CDN, and it is the only network request this page makes after it loads. Your document is not part of it — the words stay in the tab, and the model runs through WebAssembly on your own CPU. After the first download the file is cached and the feature works offline.
That is why the button is a button and not an automatic step: a tool that quietly phones a CDN the moment you paste something is technically "processing locally" while still telling a third party when you used it. If you would rather not download anything, never press it — the pattern engine works with no network at all.
The three mask styles, and when to use each
Placeholders replace each value with a stable token like <EMAIL_1>. The same email becomes the same token every time, so the document still shows that two mentions are the same party — which in a contract is information you need. Shape keeps the format (j•••@e•••.com) so a reader knows what kind of thing was there. Redact removes it entirely.
Why this runs locally when the point is to use an AI
If you are about to paste a document into an AI provider, this is the step to do first: mask it here, then send the masked copy. The provider never sees the values, and you keep the mapping yourself if you need to translate the answer back. That is the same protection people buy a seven-thousand-dollar machine to get by hosting a model at home — done with detection instead, at no cost.