Rizzo-AI-Academy/rizzo-pii

Local-first privacy guard: anonymize your documents before sharing with LLMs.

What it solves

It addresses the privacy risk of sending sensitive personal and confidential data to cloud-based LLMs (like ChatGPT, Claude, or Gemini). For professionals in law, accounting, and notary services—particularly in Italy—uploading documents often violates GDPR compliance. This project allows users to leverage powerful frontier models without transmitting real identifiable information.

How it works

The system implements a reversible anonymization workflow that runs entirely on the user's local CPU:

  1. Local Detection: A lightweight 0.3B parameter model (based on mmBERT/ModernBERT) identifies 22 categories of PII, including specific Italian legal identifiers like codice fiscale and partita IVA.
  2. Pseudonymization: Real data is replaced with stable, type-aware placeholders (e.g., [FULLNAME_1]). A mapping dictionary is stored locally on the user's disk.
  3. Cloud Processing: Only the anonymized text is sent to the frontier LLM.
  4. Local Restoration: When the LLM responds, the tool uses the local dictionary to swap placeholders back into the original real values.

To ensure high accuracy, the neural model is paired with a deterministic regex and checksum network (e.g., Luhn for cards, mod-97 for IBAN) that overrides the model when a valid checksum is found.

Who it’s for

Law firms, accountants, notaries, and healthcare providers bound by GDPR who need to use LLMs for summarizing contracts or drafting legal documents while keeping sensitive data on their own devices.

Highlights

  • Italian-Legal Specialization: Specifically detects Italian-specific identifiers (CF, PIVA, cadastral data) often missed by generic models.
  • Low Hardware Requirements: Runs on a standard laptop CPU with a memory footprint of approximately 0.5 GB to 1.2 GB.
  • Reversible Workflow: Unlike destructive redaction, it allows the final output to be restored to its original form locally.
  • Privacy by Design: Operates fully offline with no API keys or telemetry, ensuring data never leaves the device during the anonymization phase.

Related

  • Project
    data-privacy-stack/presidio

    A data protection SDK that identifies and anonymizes Personally Identifiable Information (PII) in text and images using customizable recognizers.

  • Dispatch
    OpenAI Privacy Filter: Building Scalable PII Detection Web Apps

    OpenAI has released Privacy Filter, an open-source 1.5B-parameter PII detector capable of labeling eight categories of sensitive data across a 128k context window.

  • Project
    amaiya/onprem

    A privacy-focused Python toolkit for document intelligence that enables local-by-default LLM execution on sensitive data with optional cloud integration.

  • Project
    LearningCircuit/local-deep-research

    An AI-powered research assistant that performs deep, agentic research using local LLMs and multiple search engines to generate cited reports from the web and personal documents.

  • Project
    dataiku/kiji-proxy

    Kiji Privacy Proxy is an open‑source local proxy (macOS desktop app or Linux server) that automatically detects and masks 26 types of personally‑identifiable information in requests to AI services (OpenAI, Anthropic, Gemini, etc.) using an ONNX‑optimized DistilBERT model. It replaces sensitive data with realistic dummy values before forwarding the request, then restores the original data in the response. The tool works transparently via HTTP proxy settings (PAC on macOS, env vars on Linux), includes a Chrome extension for web‑based chat, offers a UI for reviewing masked mappings, and is configurable via custom regexes. All PII processing happens locally, keeping data private and helping meet GDPR/CCPA/HIPAA requirements. Installation is via Homebrew cask on macOS or Debian/Ubuntu packages, Docker, or tarballs. The project is Apache‑2.0 licensed and maintained by Dataiku’s 575 Lab, with full documentation, contribution guides, and security contact information.