Limina Data De-identification (formerly Private AI)

Limina, which rebranded from Private AI in March 2026, was founded by privacy and machine learning experts from the University of Toronto and has its mailing address in Toronto, Canada. It helps regulated organizations use sensitive text, documents, images and audio for AI, analytics and research without exposing personal, health or payment data.

Its de-identification engine uses context-aware machine learning rather than regular expressions to detect more than 50 PII, PHI and PCI entity types in 52 languages, links mentions of the same person through coreference resolution, and then redacts, pseudonymizes, reversibly tokenizes or replaces them with synthetic values. It ships as a single container that runs in the customer's VPC or on-premise, so data does not leave the customer's environment. Users include Providence, Boehringer Ingelheim, Zurich Insurance and MUFG Bank.

• Context-aware detection: Finds 50+ PII, PHI and PCI entity types the way a trained reviewer would.
• Multilingual and multimodal: Handles text, PDFs, images, audio and DICOM in 52 languages.
• Flexible transformation: Redact, pseudonymize, reversibly tokenize or generate synthetic data.
• Coreference and relation linking: Connects names, titles and attributes of the same person.
• Custom entities: Adds regex rules and tuned models for domain-specific identifiers.
• Container deployment: Runs in the customer's VPC or on-premise with no outbound data calls.

A Vietnamese hospital can de-identify medical records before sharing them with a research partner. A bank contact center can redact customer names and account numbers from call transcripts before training a speech analytics model. A health insurer can pseudonymize claims documents for fraud analytics.

Restricted datasets can be used for AI training and analytics while meeting HIPAA Safe Harbor, GDPR and PCI DSS expectations. The vendor reports less than half the error rate of AWS Comprehend, Google DLP and Microsoft Presidio on real-world data and redaction at 70,000 words per second on GPU. Providence reached over 99.5% accuracy on target PHI entities with no data exposed to third parties. Deployment takes two Docker commands.

Hospitals, banks and insurers hold large archives of call recordings, chat logs and clinical notes that they want to use for AI, but privacy laws restrict processing of sensitive data. Vietnam's 2026 data protection rules require anonymization or encryption for sensitive data transfers, increasing demand for on-premise de-identification tools.

Private AI rebranded as Limina in March 2026; the old private-ai.com domain redirects to getlimina.ai. The privacy statement lists 366 Adelaide St W, Toronto, Ontario. ISO 27001 certified; technology independently validated by Armilla AI. Partnerships announced with NVIDIA (February 2025) and Replica Analytics (May 2025).