PrivaMesh Edge

PrivaMesh Edge is the browser-first, local/offline edition of PrivaMesh for semantic email anonymization.

It is designed for Outlook add-ins and browser workflows where raw email content should stay on the user's device. The model runs with Transformers.js, ONNX Runtime Web, and WebGPU where available, with deterministic local safety fallbacks in the application.

This repository is a PrivaMesh-branded edge model package based on Gemma ONNX weights optimized for browser inference.

Positioning

PrivaMesh Edge demonstrates the core product message:

Semantic anonymization without sending raw email content to a cloud AI endpoint.

The model is intended to help:

• preserve business and cybersecurity context;
• pseudonymize personal data, account IDs, emails, phone numbers, IBANs, IPs, CVEs, and similar sensitive identifiers;
• generate synthetic anonymized email bodies;
• run locally in the browser with WebGPU acceleration where supported.

Browser Usage

import { pipeline } from "@huggingface/transformers";

const generator = await pipeline(
  "text-generation",
  "sallani/PrivaMesh-Edge",
  {
    device: "webgpu",
    dtype: "q4",
  },
);

const messages = [
  {
    role: "system",
    content:
      "You are PrivaMesh Edge. Generate a synthetic anonymized email body. Preserve business and cybersecurity meaning. Replace all private identifiers with realistic fake values. Output only the email body.",
  },
  {
    role: "user",
    content:
      "Hello Ahmed, the audit mentions Sophie Bernard, sophie.bernard@acme-finance.fr, +33 6 72 55 18 44, IBAN FR76 3000 4000 5500 0001 2345 678, host 10.18.4.22 and CVE-2025-18422.",
  },
];

const output = await generator(messages, {
  max_new_tokens: 180,
  temperature: 0,
  do_sample: false,
});

console.log(output[0].generated_text.at(-1).content);

Outlook Add-in Integration

Set the model ID in the PrivaMesh Outlook Edge add-in:

VITE_PRIVAMESH_MODEL_ID=sallani/PrivaMesh-Edge

Relationship to PrivaMesh

• sallani/PrivaMesh is the existing PrivaMesh SLM/PII model line.
• sallani/PrivaMesh-Edge is the browser-first edition focused on local WebGPU execution.
• A larger Gemma 4B-class LoRA/ONNX edition can be trained later, but this repository provides a practical browser-ready starting point now.

Local/Offline Note

After the browser downloads and caches the model, inference can run locally without sending raw email content to a remote AI inference endpoint. First load still requires fetching model assets unless the application bundles or pre-caches them.

Limitations

• PrivaMesh Edge does not guarantee perfect anonymization or regulatory compliance by itself.
• WebGPU support depends on browser, operating system, and hardware.
• Semantic anonymization can miss structured identifiers, rare formats, ambiguous context, or domain-specific sensitive data.
• The model has not been certified for GDPR, HIPAA, SOC 2, ISO, or any other regulatory framework.
• The model should not be treated as production-ready without additional validation, monitoring, and enterprise controls.
• A local deterministic anonymization layer should remain enabled for safety and repeatability.
• This repository inherits the Gemma license obligations from the base model.

Recommended Production Architecture

For production-grade workflows, semantic anonymization should be combined with deterministic and enterprise-grade controls:

Raw email or document
    -> Deterministic detectors
       - email regex
       - phone validation
       - IBAN validation
       - URL/IP detection
    -> PrivaMesh Edge semantic anonymization
    -> Enterprise DLP / policy checks
    -> Human validation if required
    -> Sanitized text for downstream AI assistant

The deterministic layer catches structured identifiers with high precision. The semantic model helps detect contextual sensitive information that may not follow fixed patterns. The DLP and human validation layers reduce residual risk in regulated environments.

Recommended complementary controls include:

• regex-based email detection;
• phone number validation;
• IBAN detection;
• URL and IP address detection;
• enterprise DLP systems;
• human validation for high-risk or regulated workflows.

Security & Privacy

PrivaMesh Edge is designed for local-first processing: raw email content can be processed in the browser with WebGPU acceleration, without sending the text to a remote AI inference endpoint. This privacy-first architecture reduces data movement and supports offline-capable workflows after model assets are available locally or cached by the browser.

Local execution does not remove the need for security controls. Applications should still apply deterministic redaction, output inspection, audit logging where appropriate, and enterprise policy checks before sanitized content is shared with downstream assistants or external systems.

Responsible Use

Use PrivaMesh Edge as one layer in a broader privacy and security architecture. Do not rely on the model as the sole control for regulated data handling, legal compliance, or external disclosure review.

For sensitive or regulated environments, validate outputs before use, test against representative data, document residual risks, and keep a human review path for high-impact workflows.

Brand

PrivaMesh Edge is part of the PrivaMesh local-first privacy architecture.

Website: https://www.privamesh.com/ Existing PrivaMesh model: https://huggingface.co/sallani/PrivaMesh