$ cat ~/build-log/oliver-identity-systems.log
Building invisible UI on macOS
// What Oliver is teaching me about native windowing, screen-capture exclusion, and keeping AI useful during live calls.
Desktop AI has different constraints than web: OS permissions, audio routing, and screen-capture exclusion all show up in the first week.
The hardest part of Oliver was not the model integration — it was making the window truly invisible. macOS uses NSWindowSharingNone at the window level, but you still have to verify against every major conferencing app. Zoom, Teams, and Google Meet each handle screen capture differently.
Audio routing exposed another layer. Getting both mic and system audio into a single stream requires a virtual audio device like BlackHole, and the setup has to be explained clearly in onboarding or users give up before the first transcription.
Privacy became a design constraint, not a feature. Storing API keys with AES-GCM-SIV encryption and keeping all history in local SQLite means there is no backend to breach. That is a trust story I can tell a user in one sentence.
The lesson: native product development compresses your tolerance for abstraction. When you own the window, the audio graph, and the key store, every shortcut shows.