Voxnotes turns a voice memo into structured meeting notes: a summary, a list of decisions, and action items with owners. It's built as two stages with different cost profiles.
Transcription runs locally via faster-whisper on our own server CPU. No third-party account, no per-minute API billing — $0 marginal cost per request. The only real constraint is our own compute capacity, which is why usage caps are expressed in minutes as well as request count.
Extraction is a single schema-constrained LLM call: the raw transcript goes in, and a JSON schema (summary, decisions array, action_items array with task+owner) comes out. One call, no chaining. We measured the real cost against the production code path rather than estimating: average 253 input tokens and 1080 output tokens per run, which comes out to about $0.002 per extraction at published model pricing.
The failure mode we specifically designed against is fabrication, not just missed items. A summarizer under pressure to be useful will happily invent a plausible-sounding decision or action item that was never said — for a notes tool, a fabricated action item is worse than a missed one, because it's actively misleading. The extraction prompt is explicit: extract only what's actually said, and if there are no real decisions or action items, return empty arrays rather than filling the schema with invented content. Our accuracy fixtures include transcripts with nothing to extract on purpose, to check the model does the honest thing instead of the shape-filling thing.
Voxnotes is a small paid product built on this pipeline, currently free to try with no signup required.
