How a Tiny Team Ships Like a Big One · The Product Track
The standup dashboard nobody asked for
One dependency-free Python file, 140 meeting transcripts, a filler-word leaderboard, and profanity per 1,000 words - rebuilt every Monday by a scheduled agent.
Every meeting we hold leaves a transcript behind - a year of standups, reviews, and retros sitting in a folder as text. At some point, with everything else in this series up and running, it was inevitable that someone here would look at that folder and think: that's a dataset.
So, for a change of register: the least necessary thing we've built, and one of the most beloved. A dashboard about our own meetings. It has a filler-word leaderboard. It tracks profanity per 1,000 words. Nobody asked for it, which we've concluded is a feature of the genre.
One file, no dependencies, one second
The build had constraints only a side project gets: it had to be trivial, or it shouldn't exist. The whole thing is one Python file with zero dependencies outside the standard library. It parses all ~140 transcripts and renders a self-contained HTML report in about one second. No pipeline, no framework, no build step - the kind of engineering the rest of this series would frown at, deployed exactly where it belongs.
The one genuinely interesting technical problem: the transcripts don't say who's speaking - just the recorder's own microphone versus everyone else, an undifferentiated "them." Per-person stats looked impossible until we noticed our own habit: standups run round-robin, and the recorder hands off by name - "thanks, over to you" - in short, name-bearing utterances. Detecting those handoffs lets the parser infer speaking slots from conversational structure. The lesson generalizes further than the toy: your data's structure is often in its rituals, not its schema.
What ~19 sections of meeting analytics look like
Meeting-length trends (are standups creeping?). Talk-ratio balance, including the manager share - a number I watch with personal interest and occasional discomfort. A co-mention graph of whose names come up together - the informal org chart, which only partly resembles the formal one. Vocabulary diversity. A topic-of-the-month tracker, which is an accidental historical record of what each era of the company obsessed over. The filler-word leaderboard. And yes, profanity per 1,000 words, which mostly proves our incidents have been mild lately.
Team-level, all of it. The report deliberately attributes nothing to individuals - the filler-word leaderboard is funny exactly once before it becomes a surveillance tool, and we like the joke too much to let it curdle. It's a mirror for the team, not a manager's scorecard.
And the punchline this series requires: the dashboard maintains itself. Every Monday, a scheduled cloud agent clones the repo, rebuilds the report over whatever new transcripts exist, diffs the output, commits only if something changed, and messages me if anything failed. Unlike the GitHub dashboard, this one was never allowed to go stale - the toy got the automation the serious tool initially lacked, which says something true about how engineers allocate love.
Real numbers
- ~140 transcripts parsed, ~19 report sections, ~1 second end to end, zero dependencies.
- Every Monday, rebuilt by a scheduled agent - diff-aware, so no-change weeks commit nothing.
- Zero per-person stats published, by policy that predates any temptation.
- One conversational ritual (round-robin handoffs) doing the work a speaker-diarization model would have been hired for.
Where the humans sit
Laughing at the leaderboard, mostly - which is the point. But it earns one serious note: everything on this blog has been about a system where agents do the toil and people keep the judgment, and the pattern holds even here, at the silly end of the spectrum. The agent rebuilds the report; humans decide what it's for - and decided, before it ever shipped, what it would never be used for. That decision, made early and kept, is the entire difference between a team that instruments itself and a team that surveils itself.
Steal this
Your meeting transcripts are a dataset your team already owns; a culture dashboard is one file and one scheduled agent away, and the return on morale is embarrassingly high for the effort. Two rules if you try: keep every stat team-level, and let it be funny - the humor is load-bearing, because a dashboard the team laughs at is one the team actually reads. Look for your rituals in the data. They're more parseable than you think.
That's the story so far: three tracks, one thesis held throughout - toil to the agents, judgment to the people, and every gate we called non-negotiable still is. The blog keeps going; this team keeps building things worth writing about. Thanks for reading this far.
This post is part of How a Tiny Team Ships Like a Big One, an ongoing series on how six builders run a production AI company. Building at Aithon - if this is how you want to work, talk to us.