VaultOS: local AI over an encrypted vault
Personal knowledge that never leaves your machine — and can still be queried in natural language.
The problem
Useful knowledge tools all ask for the same thing: upload everything to the cloud. And when they add AI, your private material becomes context for someone else's service. The question wasn't whether that feels wrong, but whether it can be avoided without losing usefulness.
The architecture decision
Local-first, non-negotiable: layered encryption (AES-256-GCM, Argon2id, biometric unlock) and the language model running on the device. No cloud, no telemetry. The hard consequence of that decision is that the local model is small, so the system had to be designed around its limits instead of ignoring them.
The principle it proves
Privacy is designed into the architecture or it doesn't exist. A privacy policy is a promise; data that never leaves the device is a guarantee.
Current state
Functional on macOS. The journal captures the most uncomfortable finding: a small LLM does not know what it does not know.