gdpr-pseudonymizer 1.0.2

CLI tool for GDPR-compliant pseudonymization of French text documents using NLP-based entity detection and reversible mapping

GDPR Pseudonymizer

AI-Assisted Pseudonymization for French Documents with Human Verification

Transform sensitive French documents for safe AI analysis with local processing, mandatory human review, and GDPR compliance.

---

🎯 Overview

GDPR Pseudonymizer is a privacy-first CLI tool that combines AI efficiency with human accuracy to pseudonymize French text documents. Unlike fully automatic tools or cloud services, we prioritize zero false negatives and legal defensibility through mandatory validation workflows.

Perfect for:

• 🏛️ Privacy-conscious organizations needing GDPR-compliant AI analysis
• 🎓 Academic researchers with ethics board requirements
• ⚖️ Legal/HR teams requiring defensible pseudonymization
• 🤖 LLM users who want to analyze confidential documents safely

---

✨ Key Features

🔒 Privacy-First Architecture

• ✅ 100% local processing - Your data never leaves your machine
• ✅ No cloud dependencies - Works completely offline after installation
• ✅ Encrypted mapping tables - AES-256-SIV encryption with PBKDF2 key derivation (210K iterations), passphrase-protected reversible pseudonymization
• ✅ Zero telemetry - No analytics, crash reporting, or external communication

🤝 AI + Human Verification

• ✅ Hybrid detection - AI pre-detects 40-50% of entities (NLP + regex patterns)
• ✅ Mandatory validation - You review and confirm all entities (ensures 100% accuracy)
• ✅ Fast validation UI - Rich CLI interface with keyboard shortcuts, <2 min per document
• ✅ Smart workflow - Entity-by-type grouping (PERSON → ORG → LOCATION) with context display
• ✅ Entity variant grouping - Related forms ("Marie Dubois", "Pr. Dubois", "Dubois") merged into one validation item with "Also appears as:" display
• ✅ Batch actions - Confirm/reject multiple entities efficiently

📊 Batch Processing

• ✅ Consistent pseudonyms - Same entity = same pseudonym across 10-100+ documents
• ✅ Compositional matching - "Marie Dubois" → "Leia Organa", "Marie" alone → "Leia"
• ✅ Smart name handling - Title stripping ("Dr. Marie Dubois" = "Marie Dubois"), compound names ("Jean-Pierre" treated as atomic)
• ✅ Selective entity processing - --entity-types flag to filter by type (e.g., --entity-types PERSON,LOCATION)
• ✅ 50%+ time savings vs manual redaction (AI pre-detection + validation)

🎭 Themed Pseudonyms

• ✅ Readable output - Star Wars, LOTR, or generic French names
• ✅ Maintains context - LLM analysis preserves 85% document utility (validated: 4.27/5.0)
• ✅ Gender-preserving - When NER provides gender classification (PERSON entities)
• ✅ Full entity support - PERSON, LOCATION, and ORGANIZATION pseudonyms for all themes

---

🚀 Quick Start

Status: 🎉 v1.0.0 — Public Release (February 2026)

Realistic Expectations for v1.0

What v1.0 delivers:

• 🤖 AI-assisted detection - Hybrid NLP + regex detects ~40-50% of entities automatically
• ✅ Mandatory human verification - You review and confirm all entities (2-3 min per document)
• 🔒 100% accuracy guarantee - Human validation ensures zero false negatives
• ⚡ 50%+ faster than manual - Pre-detection saves time vs pure manual redaction

What v1.0 does NOT deliver:

• ❌ Fully automatic "set and forget" processing
• ❌ 85%+ AI accuracy (current: 40-50% with hybrid approach)
• ❌ Optional validation mode (validation is mandatory)

Roadmap

v1.0 (MVP - Q2 2026): AI-assisted CLI with mandatory validation

• Target: Privacy-conscious early adopters who value human oversight
• 100% local processing, encrypted mapping tables, audit trails

v1.1 (Q2-Q3 2026): Quick wins & GDPR compliance

• GDPR Right to Erasure: selective entity deletion (delete-mapping command, Article 17)
• Gender-aware pseudonym assignment for French names
• Beta feedback bug fixes and UX improvements

v2.0 (Q3-Q4 2026): GUI & broader accessibility

• Desktop GUI wrapping CLI core (drag-and-drop, visual entity review)
• Standalone executables (.exe for Windows, .app for macOS) — no Python required
• French-first UI with i18n architecture (multi-language ready)
• WCAG AA accessibility for professional/academic contexts
• Target: Non-technical users (HR, legal, compliance teams)

v3.0 (2027+): NLP accuracy & automation

• Fine-tuned French NER model (70-85% F1 target, up from 40-50%)
• Optional --no-validate flag for high-confidence workflows
• Confidence-based auto-processing (85%+ F1 target)
• Multi-language support (English, Spanish, German)

---

⚙️ Installation

See Installation Guide for detailed platform-specific instructions.

Prerequisites

• Python 3.10-3.12 (validated in CI/CD)

Install from PyPI (Recommended)

pip install gdpr-pseudonymizer

# Download spaCy French model (required - 571MB)
python -m spacy download fr_core_news_lg

# Verify installation
gdpr-pseudo --help

Install from Source (Developer)

# Clone repository
git clone https://github.com/LioChanDaYo/RGPDpseudonymizer.git
cd RGPDpseudonymizer

# Install dependencies via Poetry
pip install poetry>=1.7.0
poetry install

# Install spaCy French model
poetry run python -m spacy download fr_core_news_lg

# Verify installation
poetry run gdpr-pseudo --help

Quick Test

# Test on sample document
echo "Marie Dubois travaille à Paris pour Acme SA." > test.txt
gdpr-pseudo process test.txt

# Or specify custom output file
gdpr-pseudo process test.txt -o output.txt

Expected output: "Leia Organa travaille à Coruscant pour Rebel Alliance."

Configuration File (Optional)

Generate a config template to customize default settings:

# Generate .gdpr-pseudo.yaml template in current directory
poetry run gdpr-pseudo config --init

# View current effective configuration
poetry run gdpr-pseudo config

Example .gdpr-pseudo.yaml:

database:
  path: mappings.db

pseudonymization:
  theme: star_wars    # neutral, star_wars, lotr
  model: spacy

batch:
  workers: 4          # 1-8 (use 1 for interactive validation)
  output_dir: null

logging:
  level: INFO

Note: Passphrase is never stored in config files (security). Use GDPR_PSEUDO_PASSPHRASE env var or interactive prompt. Minimum 12 characters required (NFR12).

---

📖 Documentation

Documentation Site: https://liochandayo.github.io/RGPDpseudonymizer/

For Users:

• 📘 Installation Guide - Platform-specific installation instructions
• 📗 Usage Tutorial - Step-by-step usage tutorials
• 📕 CLI Reference - Complete command documentation
• 📕 Methodology & Academic Citation - Technical approach and GDPR compliance
• ❓ FAQ - Common questions and answers
• 🔧 Troubleshooting - Error reference and solutions

For Developers:

• 📚 API Reference - Module documentation and extension points
• 🏗️ Architecture Documentation - Technical design
• 📊 NLP Benchmark Report - NER accuracy analysis
• 📊 Performance Report - NFR performance validation results

For Stakeholders:

• 🎨 Positioning & Messaging
• 📋 Deliverables Summary

---

🔬 Technical Details

NLP Library Selection (Story 1.2 - Completed)

After comprehensive benchmarking on 25 French interview/business documents (1,855 entities):

Library	F1 Score	Precision	Recall	Decision
spaCy fr_core_news_lg	29.5%	27.0%	32.7%	✅ Selected
Stanza fr_default	11.9%	10.3%	14.1%	❌ Rejected

Why both failed 85% target:

• Pre-trained models optimized for news text (not interview/business docs)
• Domain-specific language patterns (conversational, mixed registers)
• ORG detection catastrophic (3.8% precision = 96% false positives)

Approved Solution:

• ✅ Hybrid approach (NLP + regex) targets 40-50% F1
• ✅ Mandatory validation ensures 100% final accuracy
• 📅 Fine-tuning deferred to v3.0 (70-85% F1 target, requires training data from v1.x/v2.x user validations)

See full analysis: docs/nlp-benchmark-report.md

Validation Workflow (Story 1.7 - Complete)

The validation UI provides an intuitive keyboard-driven interface for reviewing detected entities:

Features:

• ✅ Entity-by-type grouping - Review PERSON → ORG → LOCATION in logical order
• ✅ Context display - See 10 words before/after each entity with highlighting
• ✅ Confidence scores - Color-coded confidence from spaCy NER (green >80%, yellow 60-80%, red <60%)
• ✅ Keyboard shortcuts - Single-key actions: [Space] Confirm, [R] Reject, [E] Modify, [A] Add, [C] Change pseudonym
• ✅ Batch operations - Accept/reject all entities of a type at once (Shift+A/R)
• ✅ Help overlay - Press [H] for full command reference
• ✅ Performance - <2 minutes for typical 20-30 entity documents

Workflow Steps:

1. Summary screen (entity counts by type)
2. Review entities by type with context
3. Flag ambiguous entities for careful review
4. Final confirmation with summary of changes
5. Process document with validated entities

Deduplication Feature (Story 1.9): Duplicate entities grouped together - validate once, apply to all occurrences (66% time reduction for large docs)

Entity Variant Grouping (Story 4.6): Related entity forms automatically merged into single validation items. "Marie Dubois", "Pr. Dubois", and "Dubois" appear as one item with "Also appears as:" showing variant forms. Prevents Union-Find transitive bridging for ambiguous surnames shared by different people.

---

Technology Stack

Component	Technology	Version	Purpose
Runtime	Python	3.10-3.12	Validated in CI/CD (3.13+ not yet tested)
NLP Library	spaCy	3.8.0	French entity detection (fr_core_news_lg)
CLI Framework	Typer	0.9+	Command-line interface
Database	SQLite	3.35+	Local mapping table storage with WAL mode
Encryption	cryptography (AESSIV)	44.0+	AES-256-SIV encryption for sensitive fields (PBKDF2 key derivation, passphrase-protected)
ORM	SQLAlchemy	2.0+	Database abstraction and session management
Validation UI	rich	13.7+	Interactive CLI entity review
Keyboard Input	readchar	4.2+	Single-keypress capture for validation UI
Testing	pytest	7.4+	Unit & integration testing
CI/CD	GitHub Actions	N/A	Automated testing (Windows/Mac/Linux)

---

🤔 Why AI-Assisted Instead of Automatic?

Short answer: Privacy and compliance require human oversight.

Long answer:

1. GDPR defensibility - Human verification provides legal audit trail
2. Zero false negatives - AI misses entities, humans catch them (100% coverage)
3. Current NLP limitations - French models on interview/business docs: 29.5% F1 out-of-box
4. Better than alternatives:
   • ✅ vs Manual redaction: 50%+ faster (AI pre-detection)
   • ✅ vs Cloud services: 100% local processing (no data leakage)
   • ✅ vs Fully automatic tools: 100% accuracy (human verification)

User Perspective:

"I WANT human review for compliance reasons. The AI saves me time by pre-flagging entities, but I control the final decision." - Compliance Officer

---

🎯 Use Cases

1. Research Ethics Compliance

Scenario: Academic researcher with 50 interview transcripts needing IRB approval

Without GDPR Pseudonymizer:

• ❌ Manual redaction: 16-25 hours
• ❌ Destroys document coherence for analysis
• ❌ Error-prone (human fatigue)

With GDPR Pseudonymizer:

• ✅ AI pre-detection: ~30 min processing
• ✅ Human validation: ~90 min review (50 docs × ~2 min each)
• ✅ Total: 2-3 hours (85%+ time savings)
• ✅ Audit trail for ethics board

---

2. HR Document Analysis

Scenario: HR team analyzing employee feedback with ChatGPT

Without GDPR Pseudonymizer:

• ❌ Can't use ChatGPT (GDPR violation - employee names exposed)
• ❌ Manual analysis only (slow, limited insights)

With GDPR Pseudonymizer:

• ✅ Pseudonymize locally (employee names → pseudonyms)
• ✅ Send to ChatGPT safely (no personal data exposed)
• ✅ Get AI insights while staying GDPR-compliant

---

3. Legal Document Preparation

Scenario: Law firm preparing case materials for AI legal research

Without GDPR Pseudonymizer:

• ❌ Cloud pseudonymization service (third-party risk)
• ❌ Manual redaction (expensive billable hours)

With GDPR Pseudonymizer:

• ✅ 100% local processing (client confidentiality)
• ✅ Human-verified accuracy (legal defensibility)
• ✅ Reversible mappings (can de-pseudonymize if needed)

---

⚖️ GDPR Compliance

How GDPR Pseudonymizer Supports Compliance

GDPR Requirement	Implementation
Art. 25 - Data Protection by Design	Local processing, no cloud dependencies, encrypted storage
Art. 30 - Processing Records	Comprehensive audit logs (Story 2.5): operations table tracks timestamp, files processed, entity count, model version, theme, success/failure, processing time; JSON/CSV export for compliance reporting
Art. 32 - Security Measures	AES-256-SIV encryption with PBKDF2 key derivation (210,000 iterations), passphrase-protected storage, column-level encryption for sensitive fields
Art. 35 - Privacy Impact Assessment	Transparent methodology, cite-able approach for DPIA documentation
Recital 26 - Pseudonymization	Consistent pseudonym mapping, reversibility with passphrase

What Pseudonymization Means (Legally)

According to GDPR Article 4(5):

"Pseudonymization means the processing of personal data in such a manner that the personal data can no longer be attributed to a specific data subject without the use of additional information, provided that such additional information is kept separately."

GDPR Pseudonymizer approach:

• ✅ Personal data replaced: Names, locations, organizations → pseudonyms
• ✅ Separate storage: Mapping table encrypted with passphrase (separate from documents)
• ✅ Reversibility: Authorized users can de-pseudonymize with passphrase
• ⚠️ Note: Pseudonymization reduces risk but does NOT make data anonymous

Recommendation: Consult your Data Protection Officer (DPO) for specific compliance guidance.

---

🛠️ Development Status

All 4 MVP Epics Complete — v1.0.0 released February 2026.

• ✅ Epic 1: Foundation & NLP Validation (9 stories) — spaCy integration, validation UI, hybrid detection, entity deduplication
• ✅ Epic 2: Core Pseudonymization Engine (9 stories) — pseudonym libraries, encryption, audit logging, batch processing, GDPR 1:1 mapping
• ✅ Epic 3: CLI Interface & Batch Processing (7 stories) — 8 CLI commands, progress reporting, config files, parallel batch, UX polish
• ✅ Epic 4: Launch Readiness (8 stories) — LLM utility validation, cross-platform testing, documentation, NER accuracy suite, performance validation, beta feedback integration, codebase refactoring, launch preparation
• Total: 33 stories, 1077+ tests, 86%+ coverage, all quality gates green

---

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for details on:

• Bug reports and feature requests
• Development setup and code quality requirements
• PR process and commit message format

Please read our Code of Conduct before participating.

---

📧 Contact & Support

Project Lead: Lionel Deveaux - @LioChanDaYo

For questions and support:

• 💬 GitHub Discussions — General questions, use cases
• 🐛 GitHub Issues — Bug reports, feature requests
• 📖 SUPPORT.md — Full support process and self-help checklist

---

📜 License

This project is licensed under the MIT License.

---

🙏 Acknowledgments

Built with:

• spaCy - Industrial-strength NLP library
• Typer - Modern CLI framework
• rich - Beautiful CLI formatting

Inspired by:

• GDPR privacy-by-design principles
• Academic research ethics requirements
• Real-world need for safe AI document analysis

Methodology:

• Developed using BMAD-METHOD™ framework
• Interactive elicitation and multi-perspective validation

---

⚠️ Disclaimer

GDPR Pseudonymizer is a tool to assist with GDPR compliance. It does NOT provide legal advice.

Important notes:

• ⚠️ Pseudonymization reduces risk but is NOT anonymization
• ⚠️ You remain the data controller under GDPR
• ⚠️ Consult your DPO or legal counsel for compliance guidance
• ⚠️ Human validation is MANDATORY - do not skip review steps
• ⚠️ Test thoroughly before production use

v1.0 MVP limitations:

• AI detection: 40-50% baseline (not 85%+)
• Validation required for ALL documents (not optional)
• French language only (English, Spanish, etc. in future versions)
• Text formats only (.txt, .md - no PDF/DOCX in v1.0)

---

🧪 Testing

Running Tests

The project includes comprehensive unit and integration tests covering the validation workflow, NLP detection, and core functionality.

Note for Windows users: Due to known spaCy access violations on Windows (spaCy issue #12659), Windows CI runs non-spaCy tests only. Full test suite runs on Linux/macOS.

Run all tests:

poetry run pytest -v

Run only unit tests:

poetry run pytest tests/unit/ -v

Run only integration tests:

poetry run pytest tests/integration/ -v

Run accuracy validation tests (requires spaCy model):

poetry run pytest tests/accuracy/ -v -m accuracy -s

Run performance & stability tests (requires spaCy model):

# All performance tests (stability, memory, startup, stress)
poetry run pytest tests/performance/ -v -s -p no:benchmark --timeout=600

# Benchmark tests only (pytest-benchmark)
poetry run pytest tests/performance/ --benchmark-only -v -s

Run with coverage report:

poetry run pytest --cov=gdpr_pseudonymizer --cov-report=term-missing --cov-report=html

Run validation workflow integration tests specifically:

poetry run pytest tests/integration/test_validation_workflow_integration.py -v

Run quality checks:

# Code formatting check
poetry run black --check gdpr_pseudonymizer tests

# Format code automatically
poetry run black gdpr_pseudonymizer tests

# Linting check
poetry run ruff check gdpr_pseudonymizer tests

# Type checking
poetry run mypy gdpr_pseudonymizer

Run Windows-safe tests only (excludes spaCy-dependent tests):

# Run non-spaCy unit tests (follows Windows CI pattern)
poetry run pytest tests/unit/test_benchmark_nlp.py tests/unit/test_config_manager.py tests/unit/test_data_models.py tests/unit/test_file_handler.py tests/unit/test_logger.py tests/unit/test_naive_processor.py tests/unit/test_name_dictionary.py tests/unit/test_process_command.py tests/unit/test_project_config.py tests/unit/test_regex_matcher.py tests/unit/test_validation_models.py tests/unit/test_validation_stub.py -v

# Run validation workflow integration tests (Windows-safe)
poetry run pytest tests/integration/test_validation_workflow_integration.py -v

Test Coverage

• Unit tests: 777 tests covering validation models, UI components, encryption, database operations, audit logging, progress tracking, and core logic
• Integration tests: 90 tests for end-to-end workflows including validation (Story 2.0.1), encrypted database operations (Story 2.4), compositional logic, and hybrid detection
• Accuracy tests: 22 tests validating NER accuracy against 25-document ground-truth corpus (Story 4.4)
• Performance tests: 15 tests validating all NFR targets — single-document benchmarks (NFR1), batch performance (NFR2), memory profiling (NFR4), startup time (NFR5), stability/error rate (NFR6), stress testing (Story 4.5)
• Current coverage: 86%+ across all modules (100% for progress module, 91.41% for AuditRepository)
• Total tests: 1077+ tests (post-refactoring baseline)
• CI/CD: Tests run on Python 3.10-3.12 across Windows, macOS, and Linux
• Quality gates: All pass (Black, Ruff, mypy, pytest)

Key Integration Test Scenarios

The integration test suite covers:

Validation Workflow (19 tests):

• ✅ Full workflow: entity detection → summary → review → confirmation
• ✅ User actions: confirm (Space), reject (R), modify (E), add entity (A), change pseudonym (C), context cycling (X)
• ✅ State transitions: PENDING → CONFIRMED/REJECTED/MODIFIED
• ✅ Entity deduplication with grouped review
• ✅ Edge cases: empty documents, large documents (320+ entities), Ctrl+C interruption, invalid input
• ✅ Batch operations: Accept All Type (Shift+A), Reject All Type (Shift+R) with confirmation prompts
• ✅ Mock user input: Full simulation of keyboard interactions and prompts

Encrypted Database (9 tests):

• ✅ End-to-end workflow: init → open → save → query → close
• ✅ Cross-session consistency: Same passphrase retrieves same data
• ✅ Idempotency: Multiple queries return same results
• ✅ Encrypted data at rest: Sensitive fields stored encrypted in SQLite
• ✅ Compositional logic integration: Encrypted component queries
• ✅ Repository integration: All repositories (mapping, audit, metadata) work with encrypted session
• ✅ Concurrent reads: WAL mode enables multiple readers
• ✅ Database indexes: Query performance optimization verified
• ✅ Batch save rollback: Transaction integrity on errors

---

📊 Project Metrics (As of 2026-02-11)

Metric	Value	Status
Development Progress	v1.0.0 Released	✅ All 4 Epics Complete
Stories Complete	33 (Epic 1-4)	✅ All stories including 4.6.1 refactoring + 4.7 launch prep
LLM Utility (NFR10)	4.27/5.0 (85.4%)	✅ PASSED (threshold: 80%)
Installation Success (NFR3)	87.5% (7/8 platforms)	✅ PASSED (threshold: 85%)
First Pseudonymization (NFR14)	100% within 30 min	✅ PASSED (threshold: 80%)
Critical Bugs Found	1 (Story 2.8)	✅ RESOLVED - Epic 3 Unblocked
Test Corpus Size	25 docs, 1,855 entities	✅ Complete
NLP Accuracy (Baseline)	29.5% F1 (spaCy)	✅ Measured
Hybrid Accuracy (NLP+Regex)	35.3% F1 (+52.2% PERSON)	✅ Story 1.8 Complete
Final Accuracy (AI+Human)	100% (validated)	🎯 By Design
Pseudonym Libraries	3 themes (2,426 names + 240 locations + 588 orgs)	✅ Stories 2.1, 3.0, 4.6 Complete
Compositional Matching	Operational (component reuse + title stripping + compound names)	✅ Stories 2.2, 2.3 Complete
Batch Processing	Architecture validated (multiprocessing.Pool, 1.17x-2.5x speedup)	✅ Story 2.7 Complete
Encrypted Storage	AES-256-SIV with passphrase protection (PBKDF2 210K iterations)	✅ Story 2.4 Complete
Audit Logging	GDPR Article 30 compliance (operations table + JSON/CSV export)	✅ Story 2.5 Complete
Validation UI	Operational with deduplication	✅ Stories 1.7, 1.9 Complete
Validation Time	<2 min (20-30 entities), <5 min (100 entities)	✅ Targets Met
Single-Doc Performance (NFR1)	~6s mean for 3.5K words	✅ PASSED (<30s threshold, 80% headroom)
Batch Performance (NFR2)	~5 min for 50 docs	✅ PASSED (<30min threshold, 83% headroom)
Memory Usage (NFR4)	~1 GB Python-tracked peak	✅ PASSED (<8GB threshold)
CLI Startup (NFR5)	0.56s (help), 6.0s (cold start w/ model)	✅ PASSED (<5s for CLI startup)
Error Rate (NFR6)	~0% unexpected errors	✅ PASSED (<10% threshold)
Test Coverage	1077+ tests (post-refactoring baseline), 86%+ coverage	✅ All Quality Checks Pass
Quality Gates	Ruff, mypy, pytest	✅ All Pass (0 issues)
Supported Languages	French	🇫🇷 v1.0 only
Supported Formats	.txt, .md	📝 v1.0 scope

---

🔗 Quick Links

• 📘 Full PRD - Complete product requirements
• 📊 Benchmark Report - NLP accuracy analysis
• 🎨 Positioning Strategy - Marketing & messaging
• 🏗️ Architecture Docs - Technical design
• 📋 Approval Checklist - PM decision tracker

---

Last Updated: 2026-02-11 (v1.0.0 Public Release — all 4 MVP epics complete)