// DATA SHIELD
Your data never leaves. The reasoning does.
Data Shield is our own real-time anonymization layer. It sits between your data and any external AI engine: it detects sensitive entities, replaces them with semantic tokens, lets the model reason over tokens, and restores real data in your perimeter. The result: you use the most powerful AI models in the world without them seeing a single real piece of data.
// HOW THE FLOW LOOKS
Your system passes sensitive data to Data Shield. Data Shield replaces them with tokens before any send. The external AI model reasons over tokens. The response comes back to your system and tokens recover their real value, only in your perimeter.
// WHAT IT IS
A privacy layer between your data and AI models.
The most powerful AI engines today (GPT, Claude, Gemini) run on external servers. To use them, you have to send them data. For companies that handle sensitive information (clients, finances, contracts, intellectual property), sending that data directly to an external provider is not acceptable.
Data Shield solves that contradiction. It's our own software layer that sits between your system and the AI models. It detects and replaces every sensitive entity with a semantic token before the model sees it. The AI model reasons over tokens; real data never leaves your perimeter. When the model returns its response, Data Shield restores the original entities in the final result.
// HOW IT WORKS
Five steps. Real-time pipeline.
01
Detection
Named Entity Recognition (NER) plus regex patterns calibrated for Argentine identifiers: CUIT, CBU, CUIL, DNI, addresses, phones, amounts, proper names. Detection models run locally on your infrastructure. No data goes out in this step.
02
Tokenization
Each detected entity is replaced with a semantic token that preserves the type: "Juan Pérez" → "Person_1", "CUIT 20-12345678-9" → "Entity_1", "$1,250,000" → "Amount_1". The AI model can reason correctly because the type is preserved, but it has no access to real data.
03
Processing
The external AI model (GPT, Claude, Gemini, or local via Ollama/vLLM) receives only the tokenized text. It reasons, analyzes, decides. Never sees a real piece of data.
04
Restoration
The model's response returns to your perimeter. Data Shield replaces tokens with original data. The token mapping lives in volatile memory; never persisted to disk. When the session ends, the mapping is erased.
05
Audit
Every anonymization and restoration operation is logged with timestamp, user, type of entity detected and result. Exportable audit trail for regulatory compliance.
// WHAT THE MODEL SEES
What you write. What the model receives.
| What you load | What the AI model receives |
|---|---|
| Juan Pérez purchased for $1,250,000 in March | Person_1 purchased for Amount_1 in Date_1 |
| CUIT 20-12345678-9, debt of USD 45,000 | Entity_1, debt of Amount_1 |
| María González, DNI 28,456,789, lives at 1234 Corrientes Ave | Person_1, Identifier_1, lives at Address_1 |
// DETECTED PATTERNS
Calibrated for Argentine and Latin American identifiers.
AR tax identifiers
CUIT, CUIL, DNI
AR banking
CBU, CBU alias, credit cards
AR geographic
Addresses (street/number/locality/province), postal codes, phones with area code and mobiles
LATAM tax identifiers
CPF (Brazil), RUT (Chile), RFC (Mexico)
International
SSN (USA), European IBAN, passports
General
Proper names (via NER), email, amounts in pesos and dollars with local format
// OPERATION MODES
Three levels based on your risk.
Standard
Default anonymization. External AI model receives tokens. Mapping in volatile memory. Sufficient for most commercial cases.
Strict (air-gapped)
Local AI models only (Ollama, vLLM). Zero traffic to the outside. Total sovereignty. For regulated sectors (health, defense) and critical data.
Audit
Maximum traceability mode. Every operation logged with full detail. For strict regulatory compliance (HIPAA, GDPR Art. 30, certifications).
// COMPLIANCE
What Data Shield solves in regulatory terms.
GDPR
Art. 6 (legal basis), Art. 25 (privacy by design), Art. 32 (security), Art. 44-49 (international transfers, the most expensive one).
HIPAA
PHI protection before external AI processing.
LGPD (Brazil)
Arts. 7, 15, 18.
Law 25.326 (Argentina)
Consent, purpose, security.
PIPL (China)
Data localization, no data export.
CCPA/CPRA (California)
Personal data protection.
Initial conversation.
Thirty minutes to evaluate your current privacy model and see how Data Shield fits.
Let's talk