Personal data never reaches the models
Removing personal data is not a setting in Howzer. It is the first thing that happens to a message, before any model sees a single word, and the analysis is just as good without it.
By Howzer Team, Engineering ·
Here is a message, roughly as it arrives. My name is Maria Schmidt, you can reach me at [email protected] or on 0172 1234567, and I have now been waiting three weeks. Here is what every model in Howzer sees. My name is [NAME], you can reach me at [EMAIL] or on [PHONE], and I have now been waiting three weeks.
The analysis is unaffected. Frustration, urgency and the underlying cause live in the structure and the vocabulary of the sentence, not in who wrote it. Which means there is no trade-off to weigh up here, and therefore no reason to make it optional.
What gets caught
Detection combines name recognition, pattern matching and rules that look at the surrounding sentence, which is what it takes to catch a name embedded in running prose rather than sitting neatly in a signature.
- Names, whether in a signature or in the middle of a sentence.
- Email addresses and phone numbers in German and international formats.
- Postal addresses, down to street, city and postcode.
- Bank details, validated by their own check digits so a random number is not mistaken for one.
- Dates of birth, distinguished from the many other dates in a service message.
- Customer identifiers, in whatever format your organisation uses.
And none of it leaves
Because Howzer runs inside your own environment, the redaction happens there too. No cloud service is called, no external name-recognition provider is involved, and there is no moment where the unredacted message is in transit to anyone.