Three-layer PII detection and anonymization for therapy session transcripts. Supports Russian and English. Fully local by default — no data leaves the machine.
Architecture
Three detection layers run in sequence, each catching what others miss:
| Layer | Tool | Catches | Size | Speed |
|---|---|---|---|---|
| 1 | Natasha | Russian names, locations, organizations | 27 MB | instant |
| 2 | OpenAI Privacy Filter (opf) | Phones, accounts, addresses, emails | 2.8 GB | ~1.5s |
| 3 | Ollama LLM | Medications, dates, contextual IDs | 2.5-7 GB | ~10s |
Spans from all layers are merged, overlaps resolved, and a unified redacted output is produced.
Prerequisites
pip install natasha setuptools pymorphy2-dicts-ru
pip install 'opf @ git+https://github.com/openai/privacy-filter.git'
ollama pull qwen3:4b
Each layer is optional — the script gracefully skips unavailable layers and warns.
Usage
Single file
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt
Stdin pipe
cat session.txt | python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py
Batch processing
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py --batch ~/sessions/ -o ~/sessions_clean/
JSON report
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --json
Pseudonyms instead of tags
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --pseudonyms
Select layers / model
# Fast — Natasha only
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --layers natasha
# LLM only — maximum coverage
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --layers ollama --model gemma4:e2b
Encrypt output (AES-256)
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt -o clean.txt --encrypt "password"
Invoking from Claude Code
To anonymize text already in context, pipe it through the script:
echo '<text>' | python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py --json
For files, pass the path directly. Always recommend manual review after automated anonymization.
Limitations
- Contextual identifiers ("the only red-haired architect in Kostroma") are NOT detected by any automated tool
- OPF is English-focused — Russian coverage is partial
- Medications detected only by Layer 3 (requires Ollama)
- Does not assess re-identification risk from combinations of non-PII fields
Guardrails
- NEVER send raw transcripts to cloud services
- Cloud verification only on already-anonymized text
- Always recommend manual review for therapy data
- Never log original PII values