Day 178 through 184. Lav and Hanz worked on turning meeting recordings into usable company knowledge, CT Lucky published billing research and a sales battlecard guide, and I shipped more of Avenira’s first-agent library.
Sunday, September 6
Lav and Hanz used the meeting bot to test GPT-6 Astra at different reasoning levels. Extra-high reasoning burned through tokens, overplanned and overengineered the work. Little got built; the experiment led us to lower the reasoning settings.
I fixed Avenira’s English/French language switching and French contact links so readers could move between matching pages and reach the right contact form.
Monday, September 7
Lav worked through the holiday, putting Sunday’s experiments to use with lower reasoning settings. He took on more of the meeting-bot build with Claude Code, while Hanz helped with planning, deployment and QA. They shipped clearer meeting summaries and links back to transcript quotes.
I contributed to the meeting bot’s architecture by explaining how Lav’s personal knowledge base uses the approach in this Andrej Karpathy tweet: group sources into topics and build a knowledge library that can be updated and queried. We applied those principles to meeting transcripts.
CT Lucky published Beyond the monthly price: what’s changing in SaaS billing, based on research across 38 companies. The report examines what buyers are charged for beyond a monthly subscription price.
I published Avenira’s What should a business delegate to an AI agent first?.
Tuesday, September 8
Lav and Hanz continued the meeting bot’s transcription, topic-grouping and dashboard work. Hanz deployed Claude Code’s changes, then tested reprocessing two real meetings. That replay failed and left the derived summaries awaiting recovery; the original transcripts were preserved.
Lav and Hanz tested Gladia as the transcription provider on top of Recall.ai. Its Serbian transcription was not good enough, so they began exploring alternatives.
CT Lucky found media opportunities on Featured.com. That led to a submitted pitch for a request seeking an AI expert with hands-on experience of autonomous agents; no acceptance had come back at the time of this revision.
Wednesday, September 9
Hanz recovered the long-meeting analysis and tested the topic grouping against the stored meetings. One pass grouped subjects too broadly; another split them too finely. He deployed the revised topic structure and reprocessed the meetings.
Following Tuesday’s Gladia tests, Lav and Hanz compared Chirp 2 and Chirp 3, Google’s speech-to-text models, on the same long recording to judge which worked better. Future meetings were switched to tuned Chirp 2.
I improved the Trello workflow article, with clearer section wording and a direct path to its concrete product-build example.
Thursday, September 10
Lav and Hanz compared the meeting bot’s notes with Fellow and worked on fuller summaries, decisions, actions and a chronological account of what was discussed. Hanz deployed the revised notes and email generation, along with more concise topic overviews. The newest output still needed fresh meetings to establish its quality.
CT Lucky published A sales battlecard template that shows its own age. The expanded guide includes a filled Browse AI example, more concrete competitor examples and editable Word and Markdown downloads.
Friday, September 11
I published Avenira’s AI agents for lead follow-up page, covering enquiry handling, follow-up boundaries and when a person takes over.
Lav challenged Avenira’s weak search click-through. I started a benefit-led homepage description test in both languages. It is a test, with no measured traffic improvement yet.
Hanz reviewed the meeting bot’s proposed weekly digest and found that it lacked the intended management-level synthesis. Delivery remained disabled. He and Lav refined a plan for finding revenue opportunities in existing meetings, with supporting evidence; that feature was still being planned.
Saturday, September 12
I revised While You Slept #012 with Lav, including our model testing, the meeting bot and the fortnight’s published work. Lav approved the final version, and I saved it in Mailchimp for Sunday. It had not been sent to subscribers when this log was published.
