Case study: a privacy layer before the LLM

The challenge

Companies in Brazil want to use language models with customer text. This text contains CPF and CNPJ numbers, e-mail addresses, phone numbers and names.

Many of these companies must not send these values to a model provider. They need a barrier before each call.

Architecture

• The barrier replaces CPF and CNPJ numbers, e-mail addresses, phone numbers and names with markers, for example [CPF_1], before the model.
• The map from value to marker exists only in the memory of the request. Usage records keep categories and counts, never the values.
• Each organisation selects its policy: placeholder, redact, block or monitor. A request can make the policy stricter, never weaker.
• If the barrier fails, the request fails. Raw text never goes to the provider as a fallback.
• The API uses the OpenAI and Anthropic formats. Streamed answers pass through without buffering.

In production

• Site and documentation in Portuguese and English, with a 30-day trial and no card.
• Credit in Brazilian reais by Pix, and per-use billing for each paid capability.
• A call without credit can pay with x402 (USDC on Base), or it receives HTTP 402 with the price.
• The audit copy of the prompt and the answer exists only in encrypted form (AES-256-GCM).

• Rota Nacional
• API documentation

Numbers and sources

• R$ 0.002 per moderated message (public price table in the documentation, checked on 14 October 2026).
• Credit packs of R$ 10, R$ 25, R$ 50 and R$ 100 by Pix (same table).
• 4 policies for each organisation: placeholder, redact, block and monitor (Rota Nacional privacy page).
• 30 days of trial with a starter quota, with no card and no API key (documentation).

• Rota Nacional privacy page

Limits

• The barrier reads text. It cannot see images or audio: the chat refuses media, and documents go through extraction first.
• The monitor policy records personal data but does not remove it. It is the highest-risk option, and the screen says so.
• The benchmark of the barrier is not public yet. This page will link to it when it is published.

Hire me for this

• A privacy layer in front of your company's LLM calls, with a policy for each organisation and an audit trail.
• An OpenAI-compatible gateway with quotas, metered usage and per-call billing.
• A review of the personal data that reaches your providers today.