Productivity meets Control.

SOWA Privacy acts as a client-side security layer in the form of a browser extension. No own servers, no cloud components.

  • Local Real-Time Analysis

    Texts are analyzed before they even leave the user’s browser.

  • Automated Anonymization

    Detection of names, IBANs, health data, etc., and replacement with neutral placeholders.

  • Provider Independent

    Works with ChatGPT, Claude, Gemini & Co. without API modification of the target systems.

  • Seamless De-Anonymization

    Anonymized values are restored to their originals when the AI responds. You get the normal results without letting personal data reach foreign chatbot servers.

How detection works

Three layers, all running locally in your browser. None of them needs to send your text to a server before deciding whether it contains personal data.

Layer 1 – Regex

Pattern matching

Always on. Fastest. No model required.

  • Built-in pattern packs for emails, phone numbers (including German mobile), IBAN, BIC/SWIFT, PESEL, NIP, credit-card numbers, driver’s licence, passport, dates of birth, IP addresses, API keys, passwords, tokens.
  • Runs on every keystroke. Zero network, zero model load – pure local string matching.
  • User-extensible: add custom rules in Settings using a four-field format (KIND | REGEX | FLAGS | PRIORITY).
Layer 2 – NER

Named entity recognition

Multilingual. Local model. Opt-in.

  • Catches what regex can’t: people’s names, organisations, locations – anything that depends on context, not pattern.
  • Multilingual DistilBERT model (~65 MB) via Transformers.js + ONNX. Downloads once from HuggingFace, then runs offline forever.
  • Runs in an isolated offscreen document so the chat page can’t see or call it.
  • Configurable confidence threshold (default 0.7) and priority weighting.
  • Default OFF. The 65 MB download only starts after the user explicitly enables it in Settings.
Layer 3 – Local LLM

Power Mode

Edge cases. Local GPU. Opt-in.

  • Catches the contextual edge cases the first two layers miss – ambiguous entities, paraphrased identifiers, sensitive details only obvious from sentence context.
  • WebLLM runtime on the user’s GPU. Choice of Phi-3, Gemma 2B, or SmolLM2 (each ~200 MB).
  • 2–4 GB RAM usage when active. Slower than the first two layers – used as a check, not the primary scanner.
  • Default OFF. The model download only starts after the user opts in.

Detection hits from the three layers are merged and ranked by priority. On top of all three, a user-managed Blacklist (priority 100, always wins) and Whitelist let individuals or IT admins shape detection for their workflow – add company-specific identifiers, exempt safe boilerplate, ship presets across a team – without touching code.

Compare to alternatives See the source code
01

Law Firms & Legal Departments

Use AI for contract review, case summaries, legal memo drafting, NDA comparison, and client communication without exposing client-identifying information.

Sensitive data examples
Client names Opposing parties Contract numbers Case numbers Addresses Court references Confidential legal facts
02

Tax Advisory

Use AI to explain tax office letters, draft client emails, summarize tax cases, and create document checklists while protecting taxpayer data.

Sensitive data examples
Tax IDs Client names Income details Company records Addresses Bank data Tax office references
03

Accounting & Bookkeeping

Use AI for invoice summaries, payment reminder drafts, expense categorization, accounting questions, and audit preparation notes without exposing invoice or payment data.

Sensitive data examples
Vendor names Customer names IBANs Invoice numbers Tax IDs Payment amounts Internal references
04

Healthcare, Clinics & Therapists

Use AI to structure medical notes, draft patient-friendly explanations, prepare appointment follow-ups, and anonymize therapy notes before using AI tools.

Sensitive data examples
Patient names Dates of birth Insurance numbers Diagnoses Medication references Appointment data Health records
05

Banking & Financial Services

Use AI for customer complaint summaries, KYC case notes, AML documentation support, audit preparation, and customer response drafts in regulated financial workflows.

Sensitive data examples
Customer names Account numbers IBANs Transaction references Loan numbers Addresses Compliance case IDs
06

Insurance

Use AI to summarize claims, explain policies, prepare missing-document checklists, draft customer replies, and structure escalation notes while masking policyholder and claim data.

Sensitive data examples
Policyholder names Claim numbers Policy numbers Accident locations Vehicle plates Health information Damage reports
07

Public Administration

Use AI to draft citizen responses, summarize applications, prepare internal memos, explain forms, and support permit or benefit-related communication.

Sensitive data examples
Citizen names Addresses Permit numbers Application IDs Social benefit info Health or disability data Official case numbers
08

HR & Recruiting

Use AI for CV summaries, interview question generation, candidate comparison, performance review drafts, and employee communication without exposing applicant or employee identities.

Sensitive data examples
Candidate names Employee names Salaries Contact details Addresses Absences Performance notes Employee IDs
09

Consulting

Use AI to summarize workshops, draft strategies, analyze processes, prepare proposals, and structure risk assessments without leaking client-confidential information.

Sensitive data examples
Client names Project names Internal roadmaps Financial figures Strategic documents Employee names Confidential processes
10

IT, IAM & Security

Use AI to summarize IAM tickets, explain access requests, draft incident reports, prepare audit evidence, and improve user communication while masking technical identifiers.

Sensitive data examples
Usernames Employee IDs IP addresses Hostnames System names Access roles Group names Ticket IDs Incident details