PATENTED TECHNOLOGY
Data protection for LLMs · Italian patent

Use the best models. Without giving them your data.

FASTEERMASK replaces sensitive data with labels that tell the model what type of data it is, not which data it is. The model still responds well. Real data never leaves your perimeter.

Works with the models you already use. No migration, no retraining.

YOUR PERIMETER INPUT TEXT FASTEERMASK detects, replaces, reconstructs VOLATILE MEMORY data ↔ label Clear response real data restored PERIMETER EXTERNAL MODEL ChatGPT · Claude · Gemini <NAME 45> · <CITY 37> sees the data type, not the data masked response Clear data never crosses this line.

The answer stays useful

The label preserves the meaning of the data, so the model can continue to reason correctly.

Your data stays inside

Clear text never crosses the boundary to the model provider.

The mapping is deleted

The data-to-label mapping lives in volatile memory and is deleted after use.

The problem

To ask a question, you currently send data out

Whenever someone pastes a contract, medical record, or customer list into an external model, that data leaves the company. Existing solutions share the same flaw: they protect the data by weakening the answer.

Clear data reaches the provider

Names, tax IDs, IBANs, health data, project names, and product names end up in the prompt and leave the company perimeter.

Deleting data weakens the answer

Remove the data before sending and the model loses context, producing lower-quality answers.

Random replacements create confusion

Replacing data with fake values does not solve the problem: the model no longer understands what kind of information it is seeing.

The tension to resolve is exactly this: keep clear data away from the model without degrading the quality of the answer. It is the blind spot in previous methods, and it is the problem FASTEERMASK solves.
How it works

Four steps, one real example

The example below is taken from the patent text. The values are fictional; the mechanism is the one running in production.

1Detection and maskinginside your perimeter
"The employee Mario Rossi, born in Milan on January 1, 1990, successfully completed project PRJ_0."
Detected data Assigned label
Mario Rossi <NAME 45>
Milano <CITY 37>
January 1, 1990 <DATE 104>
PRJ_0 <PROJECT 12>

The mapping stays in volatile memory inside your perimeter. The number next to a name distinguishes different people in the same text.

2This is all the model receivesoutside the perimeter
"The employee <NAME 45>, born in <CITY 37> on <DATE 104>, completed project <PROJECT 12> successfully."

"Summarize the progress of <NAME 45>"

The model knows that <NAME 45> is a person and <DATE 104> is a date. It does not know who or when.

3The model respondsstill masked
"The employee <NAME 45> successfully completed project <PROJECT 12>"
4Reconstructioninside your perimeter
"The employee Mario Rossi successfully completed project PRJ_0"

The labels become real data again by reading the local mapping. The mapping is then deleted. The user reads a complete answer; the model provider has never seen real data.

The difference

Why this is more than redaction

Redacting text is easy. Redacting it while leaving the model enough context to answer well is the patented idea.

Labels carry meaning

Not a random code: a name semantically related to the replaced data, so the model knows it is handling a person, an address, or an IBAN.

Every occurrence is distinct

A unique identifier distinguishes different people in the same text, so the model does not confuse them.

Reconstruction happens outside the model

Real data is restored locally. There is no moment when the model provider can see it.

Memory is volatile

The data-to-label mapping is temporary and deleted after use: no second archive accumulates that needs protection.

You decide what to mask

Remove anything that does not need protection and add anything automatic detection missed.

More than personal data

The same mechanism protects project names, product names, and confidential concepts: trade secrets, not just personal privacy.

Detection

What it detects in your text

Regular expressions for structured formats, named-entity recognition for people and organizations, and classification for sensitive categories.

Structured data
PhoneEmailTax IDCredit cardIBANDateTimeURLIP addressCoordinates
Named entities
Full nameAddressDate of birthPlaceOrganizationProduct nameProject name
Sensitive categories
Health dataBiometric dataGenetic dataCriminal dataEthnic originReligious beliefsPolitical opinionsUnion membershipSexual orientationPayment information
Use cases

Where it matters right away

Anywhere people are already pasting company documents into an external model, with or without authorization.

Legal

Contracts and opinions

Analyze and summarize contracts without exposing counterparties, amounts, or confidential clauses.

Human resources

Candidates and employees

Screen CVs and personnel documents without exposing identifying data.

Healthcare

Clinical documentation

Report on and summarize clinical text while keeping the patient unidentifiable.

Finance

Claims and case files

Process case files with IBANs, cards, and identity data protected at the point of entry.

Customer support

Tickets and conversations

Draft ticket responses without sending customer data outside.

R&D and product

Confidential projects

Project names, unannounced products, and internal concepts stay protected.

Integration

It sits in the middle, not in place of anything

FASTEERMASK is a layer between your applications and the model. It does not replace the model or change your stack.

The models you already use

Works with leading commercial models and open models installed in-house.

Your applications

Fits between the applications and agents already querying models, without rewriting them.

Where you need it

European cloud or installation in your infrastructure, with data and memory inside your boundary.

On the FASTEER marketplace

Available as a module of the FASTEER platform, activated inside the customer instance.

Your rules

Decide which categories to always mask, leave to the user, or ignore.

Traceable

What was masked and when remains recorded: the evidence you need when someone asks for an audit trail.

See the module on fasteer.ai
Compliance

What auditors ask for

This is not a badge: it is the fact that data does not leave, documented and demonstrable.

GDPR

Minimization applied at the right point: personal data is not disclosed to the model provider.

AI Act

A verifiable technical control over the data feeding the AI systems you use.

Trade secrets

Project names, products, and confidential concepts protected in the same way as personal data.

Intellectual property

The method is patented

This is not a feature built on top of a model: it is a registered method covering the full path of masking, model interaction, and reconstruction. The claims explicitly cite Large Language Models.

FASTEERMASK is a product based on AidaMask technology by Aida S.r.l., a company of the FoolFarm S.p.A. group. The patented technology masks sensitive data before it reaches external models. Developed in Italy, it is protected by Italian industrial invention patent no. 102023000007785, owned by FoolFarm S.p.A.

TitleMethod for protecting a user's personal data for a chatbot and related electronic system
Number102023000007785
OfficeUIBM · Italy
OwnerFoolFarm S.p.A.
Questions

What people ask first

Does the model's answer get worse?

That is the point: no. The label tells the model what type of data it is, so reasoning stays correct. That is exactly what distinguishes this method from deletion or random substitution.

Where does real data go?

It stays where it is. The data-to-label mapping lives in volatile memory inside your perimeter and is deleted after use.

Do we need to change models?

No. FASTEERMASK sits between your applications and the model you already use, commercial or open, in the cloud or installed in-house.

What if detection gets it wrong?

The detected-data list can be edited before sending: remove what does not need protection and add what was missed.

Does it only work on personal data?

No. The same mechanism covers project names, product names, and internal concepts, which are the most concrete risk for many companies.

How do we buy it?

As a module of the FASTEER platform, or as a component integrated into your architecture. We start with one of your use cases and show it on your documents.

First step

Bring us one of your documents

Tell us what your teams work with today. We will show the same question with and without masking, using your material instead of a slide.

  • A demonstration on your documents, not a presentation
  • An integration check with the models and applications you already use
  • A comparison of masked and clear responses before any commitment
Enter your name
Enter a valid business email address
Enter your organization
Select your role
Select your industry
Select an option
Optional, but useful: it helps us prepare something relevant rather than generic.

Every request is read and answered by a person. No newsletter, no automated sequence.

Request received

Thank you. We will be in touch shortly with a time proposal and a note on what we will show you.