PIIGate — Sensitive-data proxy for LLM prompts and traces

Redact PII before prompts leave your machine, restore it on the way back

Problem

Teams feeding customer data, contracts or logs into LLM prompts worry about PII leaking to third parties or sitting in traces. Existing redaction is either hard-coded in the app or discovered too late in plaintext traces. Solo developers avoid it because enterprise DLP is expensive and hard to integrate.

Solution

A local-first HTTP proxy that detects and replaces names, emails, phones, national IDs and card numbers with placeholders before forwarding to OpenAI/Anthropic endpoints, then restores them on the response. Trace logs only keep the redacted content. Extensible rules, allowlists, and an in-memory mapping table.

Users

Indie developers and small teams putting LLMs into customer workflows, SaaS handling contracts and tickets, developers in data-sensitivity-sensitive industries.

MVP

Start a local proxy and swap base_url; built-in name/email/phone/ID/card detection and placeholder replacement; automatic restore on response; trace logs with only redacted content; a YAML config for custom rules and allowlist.

Revenue

Open-source core plus a Pro subscription for multi-user, audit export, hit statistics and team rule libraries; or an enterprise per-API-call plan.

Why now

LLMs are entering production workflows, and privacy review and data-transfer compliance are becoming procurement gates. Small teams can't afford enterprise DLP, but a local proxy can close the gap.