Meetings produce decisions and then lose them. This turns a recording into a structured record: what was agreed, what was decided, and who owes what.
Recordings lose nothing and find nothing
The note-taker in a meeting is participating less than everyone else, and their notes carry their reading of what mattered. Recordings solve the wrong half of this - nothing is lost, and nothing is findable either, because searching an hour of audio for the moment somebody committed to something means listening to an hour of audio. So the decisions stay in the recording, the commitments get remembered differently by different people, and the argument that follows is about what was said rather than about what to do.
How it works
Whisper handles transcription, and the summarisation runs through the OpenAI Completion API against a fixed extraction task: pull out the agreements, the decisions, and the action items, and generate a structured TODO list from the discussion. The model choice stays open at the API rather than baked in, which is what lets the pipeline outlive a model generation without a rewrite.
Extraction is not summarisation
The distinction is the whole value. A summary is a shorter version of the meeting and it is still prose, so it is still unsearchable in the way that matters - you cannot ask it who owes what by when. Naming the categories up front, and pulling into those categories, produces something with fields. A commitment that lives in a field can be tracked, assigned, and checked off. The same commitment inside a well-written paragraph cannot.
The export is not an afterthought
PDF and Markdown, because the two audiences want different things: one wants a document to circulate and file, the other wants text that lands in a repository or a task tracker without being retyped. A record that requires retyping to be used will not be used.
Why not summarise it
- Summarise the transcript
- A summary is a new text with no guaranteed relationship to what was said. It reads well and it cannot be queried: there is no field to ask who owes what by when, and no way to check a line against the moment it came from. Extraction fills named fields and can be wrong in a way you can see.
- Search the transcript instead of extracting from it
- A transcript makes the words findable, which is not the same as making the meeting findable. You search for words; what you need is a commitment, and the sentence that carried it was probably "yeah, I can take that" - no keyword in it, nothing to match on. Extraction into named fields is what turns it into something you can ask a question of.
The unglamorous half everyone skips
Most of what an organisation decides is decided out loud, and almost none of it is written down in a form anything can query. Turning speech into fields is the unglamorous first half of every knowledge base worth having, and it is usually the half that gets skipped in favour of the retrieval.
What it proves
Turns unstructured recordings into a structured, searchable record, pulling out agreements, decisions, and action items automatically.
Category
Development
Built with
Python · Whisper AI · OpenAI API · PDF Generation · Markdown · Audio Processing
Need something similar?
The cheapest way in is two weeks. The first days work out which task would pay for itself in your processes; the rest builds that agent on your own data and measures it.