Primnox desktop AI scrubs PII locally using a DeBERTa NER model

The developer has released Primnox, a desktop AI application that processes user data locally to prevent personally identifiable information (PII) from reaching cloud servers. Before any message leaves the machine, an on-device DeBERTa NER model identifies names, emails, addresses, and phone numbers, replacing them with stable placeholders like FIRSTNAME or EMAIL.

• The system sends only the scrubbed tokens to the cloud for processing and rehydrates the real data in the returned reply.
• It includes a self-building knowledge graph derived from notes and conversations, which automatically generated 43 nodes and 184 connections without configuration.
• A deep research mode accesses 34 sources and reads full pages to produce cited reports, taking approximately 35 seconds in standard mode.
• Additional features include markdown notes with AI actions, calendar management, reminders, meeting recordings, and a Dynamic Island overlay for ambient visibility.

The software is licensed under BSL 1.1 until 2029 and is available via GitHub.