Proof point → Of the three repos in this venture, I shipped the one that worked and left the other two empty. Knowing which is which is the skill.
🩻 Problem
Meetings produce audio; decisions need text. Teams record calls into Dropbox and the recordings die there - unsummarized, unsearchable, unactioned.
🔨 Solution
PromptMaster Meeting Notes AI - an unattended pipeline:
Architecture Overview
- Folder watching - polls configured Dropbox folders on an interval, per-tenant via a
users.config.json(multi-user from day one). - Correct OAuth - a dedicated refresh-token flow helper, not a hardcoded token that dies in an hour.
- API-limit engineering - ffmpeg-based audio chunking for recordings over OpenAI’s 25 MB limit.
- Deployable - Dockerfile, docker-compose, start script, documented env setup.
The honest part
Two sibling repos - network-commerce backends framed around financial inclusion - were named, licensed, and never built. They are placeholders, kept as a record of what was deliberately deprioritized while the working utility shipped.
📜 Philosophy
Validate the AI utility first. Handle the unglamorous edges (token refresh, file limits, multi-tenancy) properly, because that’s where unattended systems die - and be honest about what didn’t get built.
🎓 Key learnings
- Production-shaped OpenAI audio pipelines: polling architectures, chunking, packaging.
- The measurable gap between naming a venture and shipping one - firsthand.
📈 Output & impact
- A complete, deployable meeting-summarization system with documented setup.
🌍 Why this matters
Original Tools & IP · The Founder Journey. Boards, teams, and programs all generate meeting audio that should become records automatically - this pipeline is that, deployable today. And the shelved-backends honesty is the judgment every organization needs: kill weak workstreams early instead of funding them out of pride.
🚀 Hire me
Want your meetings turned into records automatically? Let’s talk → · See also: The Experiments Graveyard · The thesis