Show HN: A Local-First, Reversible PII Scrubber for AI Workflows

Why It Matters

The library lets enterprises comply with strict data‑privacy regulations while preserving translation quality, a critical need for multilingual customer support and documentation. Its on‑device design eliminates reliance on third‑party APIs, reducing security risk and latency.

Key Takeaways

• Local-first reversible PII masking for translation pipelines
• Hybrid regex + ONNX NER detection balances speed and accuracy
• Semantic enrichment preserves gender and location context
• Encrypted mapping ensures PII never leaves device

Pulse Analysis

The rise of AI‑driven translation services has created a paradox for companies handling sensitive user data: sending raw text to cloud APIs violates privacy laws, yet redacting personally identifiable information (PII) degrades linguistic fidelity. Bridge Anonymization resolves this tension by performing detection, masking, and rehydration entirely on‑device. Its hybrid approach leverages lightweight regular expressions for deterministic patterns such as IBANs and credit cards, while a quantized ONNX‑runtime NER model captures softer entities like names and locations, delivering near‑full accuracy with a fraction of the memory footprint.

Beyond simple redaction, the library introduces semantic enrichment to retain crucial grammatical cues. By consulting gender‑guessing databases and GeoNames, masked tokens are annotated with attributes like gender or city versus country, allowing downstream translation engines to apply correct agreement rules in languages with rich morphology. This context‑aware masking maintains the natural flow of sentences, preventing the gender or case errors that typically arise from generic placeholders. The design choice to use static lookup tables keeps the runtime lean, while future versions may incorporate custom ML models for broader coverage.

Security is baked into every stage of the workflow. The mapping between placeholder IDs and original PII is encrypted with AES‑256‑GCM, and the library’s fuzzy tag matcher tolerates formatting variations introduced by LLM or MT outputs, ensuring reliable rehydration. Because the entire process runs locally—in Node.js, Bun, or even browsers via onnxruntime‑web—organizations can meet GDPR and other privacy mandates without sacrificing translation quality or incurring additional latency. Bridge Anonymization thus offers a pragmatic, compliant pathway for scaling multilingual AI applications.