Rolling your own meeting notes when the vendors are a no

Sometimes the honest answer to every vendor on the shortlist is no. The contract forbids sub-processors, the sector's regulated past the point where a cloud tool clears the bar, the client simply won't have their voice on someone else's disks. When that wall goes up, you can still have meeting notes. You just have to build them yourself, and I have, twice over now.

This is the escape hatch sitting behind the compliance conversation in where your meeting audio goes: once the SaaS options are ruled out, you roll your own.

the parts are deliberately boring

The stack's unremarkable, which is the point. whisper.cpp or faster-whisper turns audio into text, a local model under Ollama turns the transcript into a summary, and structured output pins the result into a fixed shape so it doesn't drift from one call to the next. This isn't new ground for me; it's my existing fully local meeting pipeline widened out past my own calls, and the structured-output discipline is what keeps the summaries machine-readable instead of chatty prose. I feed the model the transcript in twenty-minute windows and ask for a fixed set of fields: decisions, owners, dates, and open questions.

For batch work over a backlog I lean on the 3090 in the closet rather than the Mac.

faster-whisper call.wav --model large-v3 --output_format json

An hour of audio lands in roughly nine minutes on the M2 Ultra, quicker in bulk on the GPU box, and none of it touches a network.

The build buys you one thing the vendors never can: an audio path with no upload step for anyone to audit, because there's no upload.

the bill comes due in maintenance

Custody costs you, just in a different currency. The setup's genuine engineer work, the audio routing and the model files and a folder of scripts held together by one person's habits. Diarization, the who-said-what labeling, is mediocre everywhere, and my offline rig's no exception: two-person calls come out usable, five people with crosstalk and the labels smear until wrong attribution reads worse than none. There's no calendar or CRM glue, which is exactly the convenience the commercial tools are selling. And you're the maintenance team, so a macOS update can quietly break the whole chain on a Tuesday, and the pager is you.

who should actually build this

Not most people, and I want to be blunt about that, because the genre oversells itself. If a commercial on-device tool clears your compliance bar, use it and keep your weekends. I once talked a client into the homemade rig when an off-the-shelf on-device notetaker would've passed their review comfortably, and I spent a weekend plus two evenings building something they didn't need. That one's on me.

The people who genuinely should build are the ones for whom a sub-processor is a dealbreaker in writing: regulated clients, NDA-heavy work, the sectors where the question of where the audio goes has exactly one acceptable answer, which is nowhere. I keep mine running because half my calls sit under NDAs older than notetaker bots, and a pipeline with no upload step is the only thing that lets me answer for the data with a straight face. For a normal threat model it's overkill, and overkill you maintain forever is worse than a vendor you vetted once.