← Back to apps

04 / AI agents / Meeting operations

LiveKaizenAI

An AI notetaker that turns a meeting into owned action items, decisions and risks, each traced to what was said.

Live appLIVEKAIZENAI / PROTOTYPE
3ways in · owned actions out · every item quoted

LiveKaizenAI

A prototype, hosted live: explore it here.

Open in new tab

Meetings end. The follow-through doesn’t start.

Decisions get made and tasks get agreed out loud, then nobody writes down who owns what or by when. A week later the team remembers different versions of the same call.

From audio to a reviewed list of actions.

The app takes a recording, a transcript or a live call and turns it into structured output in a fixed order.

The pipeline
StepWhat it does
Transcribefaster-whisper turns audio, from an upload, a recording link or a live stream, into a timestamped transcript.
ExtractOne language-model call pulls out the summary, decisions, action items with owner, due date and priority, risks and drafted follow-ups.
TraceEvery action item carries a verbatim quote from the transcript and every decision its timestamp. An owner or date is filled in only when it was actually said.
ReviewA person checks the result, then chooses what to push to Slack or add to Google Calendar.

Paste it, stream it, or send a bot.

The same extraction runs whichever way the meeting arrives.

Input modes and outputs
ModeWhat happens
Upload / pasteA pasted transcript, an audio or video file, or a recording link. File and link audio goes through faster-whisper.
Live meetingTranscribes a shared browser tab or the microphone as the call runs, and confirms decisions, actions and risks as they come up.
Send botA notetaker bot joins a Google Meet, Teams, Zoom or Skype link once admitted, and posts a Slack report when the meeting ends.
What comes out
OutputDetail
Summary, decisions, risksEach decision carries its timestamp; risks are ranked high, medium or low.
Action itemsOwner, due date and priority, only when they were actually said, each with a verbatim quote.
Follow-upsDrafted messages and artifacts, grouped by workflow, for a person to review.
Slack and calendarOne Slack thread per workflow, plus reminders as an .ics download or in Google Calendar.

It drafts. A person decides.

The extraction is instructed never to invent an owner or a date: if it wasn’t said, the field stays empty. Nothing is sent or created automatically, so every message and reminder is approved first.

The transcription layer.

My part: set up and configured faster-whisper, the speech-to-text step that turns recordings and live audio into the transcript the rest of the app works from, and pushed it to the project repository.

A team prototype: the transcription step is the part I worked on.