Transformer models
Two language models trained to recognise personal data understand context: they tell "Rosa" the name from "rosa" the flower, or a surname that is also a city.
Because recognising personal data well in Italian text takes several models working together, and those models need resources an office computer or a browser does not have. Here is what is behind it, and how we protect the data meanwhile.
Two language models trained to recognise personal data understand context: they tell "Rosa" the name from "rosa" the flower, or a surname that is also a city.
A second language engine finds people, places and organisations using the surrounding words ("the claimant", "born in").
Tax codes, VAT numbers and IBANs are checked with their check character or digits, so a sequence that merely looks like a tax code is not mistaken for a real one.
Scanned PDFs and images are read with optical character recognition before analysis.
Application, engine and database run in a data centre in the European Union.
Files are processed in RAM and deleted as soon as we return the result; the text you send is not kept.
We do not use your texts or documents to train models: we do not keep them, so there is no archive to train on.
In tag mode the mapping between placeholder and real value is encrypted and tied to your account: only you, with the right permissions, can restore the original data.
The original text reaches our servers to be analysed: protection applies to what you then send to the AI. Automatic detection reduces exposure but can miss items, so review the result before using it.
The engine is open source (Apache 2.0) and can be installed on your own servers. It needs a server with adequate resources, not an office PC.
Examples, risks and ready-made settings for people who handle client data every day.
The Playground needs no account. The free account needs no credit card. The free account also includes 14 days of the Professionista plan, no card.