NameScrub
NameScrub removes personal details from text. You paste text in, the tool marks what it has found, you check the marks and copy the result out. Everything runs in the browser; the text does not leave your device.
What it is for
A letter is to be filed, a report passed on, minutes analysed by an AI assistant. First the names and contact details have to go. Doing that by hand is laborious, and things get missed.
There are cloud services for this, which would require you to entrust them with exactly the data you want to remove. And there are specialist applications that want installing and configuring. For the one document that needs passing on now, neither fits.
What it recognises
Always active: names, email addresses, telephone numbers and IBANs, the last of these with checksum validation. Dates and addresses can be switched on.
There is an overlap with Datenlotse on the classics: names, addresses, contact details, dates, IBANs. The cut is still a different one. Datenlotse follows an administrative taxonomy with file references, payment references and insurance numbers. NameScrub is built for general correspondence, with interactive review. For official notices take the administrative tool; for quickly cleaning something before passing it on, take this one.
How it is used
Paste the text, run the recognition, work through the findings, copy out the result. The review is not a formality: you see every mark and decide whether it stays.
Anyone with higher demands on recognition can use NameScrub+, a desktop application with a trained model for German named entity recognition. It also picks up places and organisations. The price is an installation.
What it does not do
Recognition in the browser works through a dictionary and context, not a model. An unfamiliar surname is found when a form of address precedes it, or a first name, a fitting verb, a pairing, or a repetition in the text. Standing bare at the start of a sentence, it stays.
Surnames that are also ordinary words are the hard case. “Frau Stark” is recognised, “die Aussage Stark” is not, reliably. Conversely, with names such as Richter, Koch or Bauer the tool marks too much rather than too little. That is a decision, not a shortcoming: a person missed weighs more than a word needlessly redacted.
Not recognised: file references, tax identification numbers and vehicle registrations. Dates only in numeric form, addresses only with a street suffix. The web version does not recognise places or company names at all; that would need a lexicon of proper nouns, and NameScrub+ provides one.
And no indirect descriptions: “the claimant's older brother” stays in, although the detail identifies. Review therefore remains obligatory. The tool takes the mechanical work off you, not the responsibility.
Where it comes from
Between cloud services and installed specialist applications there is a gap. The one would require you to entrust it with precisely the data you want to protect. The other is not worth it for a single document. NameScrub closes that gap: open it, paste, check, copy, all on your own device. The desktop variant came later, for higher demands on recognition.
The source code is public on GitHub, under an MIT licence.