An AI that hears the conversation, sees the screen and follows up.

Verbatim began as a private meeting recorder on my Mac. I added screen capture and live call coaching, so AI can combine what's being said with what's on screen and guide me while the call is happening. DayLog carries it through the rest of my working day.

Client calls move fast. Cloud note-takers send every word off the machine, can't tell two voices apart on a speakerphone, and only help once the call is over.

I designed it, directed the build with Claude Code and use it on every call. About 119 hours on Verbatim so far.

• Before: it starts from my calendar and records both sides of the call as separate tracks, my microphone and the computer's audio.
• During: each line is transcribed on the Mac and labelled by who said it, with a voiceprint check for speakerphone calls. Screen captures are stamped on the same clock, so "as you can see here" is kept with what was on screen.
• Live coaching: the transcript streams to a Claude session that reads it alongside the screen and suggests what to say next, the question I've been asked, the number to hold to. It responds as fast as the model can reason.
• After: a summary, action items with owners and a draft agenda for the next call, emailed when the call ends.
• DayLog: my phone records my working day. The Mac transcribes it locally, recognises my voice, writes a daily report and sends each thing I said I'd do to the right project.

• Transcription runs on my own Mac. Audio never goes to a cloud transcription service.
• Everyone on a recorded call knows it's being recorded.
• Coaching only suggests. I decide what to say.
• Summaries and coaching run on Claude. Speaker labels come from the audio track and a voiceprint, not from guesswork.
• Every recording reports how much audio it kept, and a silent microphone raises an alert during the call.

88

meetings recorded, transcribed and summarised

Source: Verbatim database, 1 Oct 2026

88.6 h

of conversation transcribed

Source: Verbatim database, 1 Oct 2026

11,191

screen captures kept alongside 70 meetings

Source: Verbatim database, 1 Oct 2026

731,301

words transcribed by DayLog across 54 days, 365 hours of audio

Source: DayLog archive, 1 Oct 2026

94% and 100%

of my own lines kept on two speakerphone calls, with 2% and 0% of the other speaker's echo let through

Source: labelled test on two recorded calls, 30 Sep 2026

270

automated tests

Source: test suite, 30 Sep 2026

Python, FastAPI, SQLite, faster-whisper, SpeechBrain ECAPA-TDNN, React, Claude, Kotlin (Android), MLX Whisper