The privacy and control layer for AI systems.
Privaro Proxy intercepts data before it reaches AI models, detects sensitive information, applies policies, and ensures that no raw PII is exposed during inference.
Use AI without exposing regulated data.
Privaro is a runtime data governance layer for AI systems and agents.
It sits between your application (or agents) and any LLM provider:
Your App / Agents → Privaro → LLM (OpenAI, Anthropic, etc.)
Privaro ensures that:
- Sensitive data is detected and transformed before inference
- Models never see raw PII
- Every interaction is auditable and certifiable
Most AI stacks today:
- Send raw data directly to LLMs
- Rely on contracts and policies for compliance
- Lack structural control of data flows
Privaro introduces:
👉 Data minimization by architecture
-
🔍 PII Detection
Hybrid detection (regex + NLP) for names, IDs, IBANs, emails, etc. -
🔐 Tokenization / Anonymization
Replace sensitive data with reversible tokens -
🧠 Policy Engine
Apply rules by pipeline, use case, or industry -
🔁 Re-identification
Recover original values safely outside the model -
📊 Audit Logs
Full traceability of every interaction -
⛓️ Certified Evidence (iBS)
Blockchain-backed verification of interactions
- Intercept input (prompt, document, agent message)
- Detect sensitive data
- Replace with tokens
- Forward sanitized input to LLM
- (Optional) Re-identify output
- Log and certify the interaction
Request access at: https://privaro.ai
curl -X POST "https://api.privaro.ai/v1/proxy/protect" \
-H "Content-Type: application/json" \
-H "X-Privaro-Key: YOUR_API_KEY" \
-d '{
"prompt": "My name is John Doe and my IBAN is ES7620770024003102575766",
"pipeline_id": "your_pipeline_id"
}'
{
"protected_prompt": "My name is [NM-0001] and my IBAN is [BK-0001]",
"detections": [
{"type": "full_name"},
{"type": "iban"}
],
"audit_log_id": "..."
}
Privaro is designed for modern AI systems, including multi-step agents.
POST /v1/agent/run/start
POST /v1/agent/protect
POST /v1/agent/reveal
POST /v1/agent/run/end
Privaro Proxy is built as:
- FastAPI backend
- Stateless processing layer
- Secure token vault
- Async certification pipeline (iBS)
Key concepts:
- pipeline_id → defines policies
- conversation_id / agent_run_id → token scope
- audit_log_id → traceability unit
- Tokenization is reversible but controlled
- Original data never leaves your trusted boundary
- Encryption: AES-256-GCM
- Optional: BYOK (Bring Your Own Key)
- Text prompts
- Emails
- Documents (PDF, DOCX, CSV, Excel)
- Agent messages
- Tool outputs
Privaro works with any LLM:
- OpenAI
- Anthropic
- Azure OpenAI
- Local models
- Legal document analysis
- Financial workflows (KYC, IBAN, AML)
- Healthcare data processing
- AI agents with sensitive data
- Customer support automation
Privaro is:
- Not a DLP tool
- Not a policy layer
- Not just observability
👉 It is a runtime control layer for AI
uvicorn app.main:app --reload
- Python 3.10+
- FastAPI
- Supabase (for persistence)
- External LLM provider (optional)
We welcome contributions.
Please open an issue or submit a PR.
MIT (or specify your license)
- Python SDK → https://github.com/Maperez1972/privaro-sdk-python
- JS SDK → https://github.com/Maperez1972/privaro-sdk-js
- Frontend → https://github.com/Maperez1972/privaro-7938a3bd
- Website → https://privaro.ai
- Docs → (coming soon)
The question is no longer:
“Can AI work?”
It is:
👉 “Can it be trusted with real data?”
Privaro makes that possible.