Lab project
Sample AI Lab Project
What was the hardest part?
Getting the model to *stop* summarising and instead extract structured commitments. The prompt is 80% of the work.
What did you learn?
How much of "agent-y" behaviour is just a good schema. Once the output shape was right, the model behaved.
What problem does it solve?
Meeting notes pile up in ten places and nobody re-reads them. This tool pulls them into one thread and surfaces only the decisions and next steps.
What's the tech stack?
A tiny Bun server, an OpenAI function-calling loop, and a Supabase table for persistence. No frontend framework — just a shareable link.
What problem does it solve?
Meeting notes pile up in ten places and nobody re-reads them. This tool pulls them into one thread and surfaces only the decisions and next steps.
What's the tech stack?
A tiny Bun server, an OpenAI function-calling loop, and a Supabase table for persistence. No frontend framework — just a shareable link.
What was the hardest part?
Getting the model to *stop* summarising and instead extract structured commitments. The prompt is 80% of the work.
What did you learn?
How much of "agent-y" behaviour is just a good schema. Once the output shape was right, the model behaved.
A tiny agent that turns messy meeting notes into crisp action items — the kind of side project you build in a weekend and keep improving on flights.
How long did it take?
About a weekend of building plus two evenings of prompt tuning.
What's next for it?
A Chrome extension that watches Google Meet captions in real time.
3× fewer forgotten action items across a month of dogfooding.