PII redaction layer for LLM applications
Consulting to remove CPF and CNPJ numbers, e-mail addresses, phone numbers and names from text before any model, with a policy for each organisation, encrypted audit and an OpenAI-compatible API.
The challenge
Prompts carry customer data. A support ticket, a contract or a call transcript can contain identity numbers, contacts and names.
Each model provider that receives this text becomes one more place where the data lives.
Architecture
- Detection before every call: rules for identity numbers, e-mail addresses and phone numbers, and a model for names.
- Markers replace the values. The map from marker to value exists only in memory, during the request.
- A policy for each organisation: placeholder, redact, block or monitor.
- Fail closed: if detection fails, the request fails.
- Audit copies only in encrypted form. Usage records keep categories and counts.
In production
Rota Nacional runs this layer in production as an API compatible with OpenAI and Anthropic.
Limits
- Detection reduces the risk, but it is not perfect. Unusual names and formats need tests with your data.
- The layer reads text. Images and audio need extraction first.
- The monitor policy keeps personal data in the prompt. It is for measurement, not for protection.
Hire me for this
- A privacy layer in front of your company's LLM calls.
- Detection tests with your documents, with the hit rate measured.
- Policies and an audit trail that your compliance team can review.
Let's assess your project
I reply with scope and a proposal.