WitnessAI

The constraint

The useful version had to disappear into the work. It could not add a bot to the call, depend on one conferencing platform, or turn every recording into a new cloud copy.

That pushed the architecture toward the Mac itself. WitnessAI taps system audio and the microphone as two distinct tracks. The split matters: “me” versus “the call” is known at capture time instead of guessed later.

What the app does

• Detects a call and captures both sides without joining it.
• Runs transcription locally with whisper.cpp.
• Separates people in the remote track and can associate a voice with a person.
• Files the title, transcript, decisions, and action items into a plain local folder.
• Reads upcoming calendar context and prepares a short brief before the next meeting.

The current interface

The floating HUD now has three native surfaces: Recording, Notes, and Clock. Recording keeps the capture state obvious. Notes gives the live conversation somewhere useful to land. Clock lets me assign the surrounding work to a project without opening a separate time tracker.

The screenshots on this page come from the current SwiftUI build. They were rendered against isolated synthetic state with automatic recording and meeting detection disabled, so no real meetings, notes, names, or clock history appear in the public asset.

The browser interaction below is a product tour, not a web recorder. It uses scripted demo content, asks for no microphone or screen-sharing permission, and sends nothing anywhere.

The privacy boundary

Audio capture, transcription, speaker separation, and voice identification happen on the Mac.

Recording consent varies by jurisdiction. WitnessAI keeps a visible recording indicator, and users remain responsible for notifying participants and following applicable law.

Summaries are a separate, explicit choice. They can run through a local Ollama model, an already-configured coding-agent CLI, an API key, or not run at all. If no model is configured, the recording and transcript still complete. The app should make that boundary visible instead of hiding it behind a generic “private AI” claim.

The reliability lesson

The hardest failure was silent.

macOS can deny system-audio capture while the recorder continues to return a valid-looking, empty track. For weeks, the app could record my microphone perfectly and report success while the other side was flat.

The installer now proves capture instead of trusting permission state: it plays a tone, records the result, and checks that the recorder actually heard it. A diagnostic command can repeat that test at any time. That small proof step matters more than another setup screen.

Current state

The app is in active development and used against real meetings on my own Mac. The next pass is about reliability, packaging, and a public download that does not ask someone else to reproduce my development environment.

The code and download stay private until that pass is ready. The case study is public now because the architecture and the mistakes are already useful.