How it started
A week of calls, and the notes gave out.
Patum began as something I made for myself, in a new job. This is how it went from there to an app, and why it works the way it does.
A new job
I had started at an established company, and the calls began on the first day, in Google Meet and in Slack huddles. Years of decisions stood behind every one of them, and most were written down somewhere in Confluence. The pages were there, and the time to read them wasn't. I wanted to do the job well, so I read in the evenings and still walked into calls a step behind.
Shelves
What helped first was Obsidian. I pulled everything that was already written into a vault and had local models sort it onto shelves: a folder for each project, each client and each person. Having one place to look changed the mornings. A question from a call could be answered that evening, from the shelf.
Then came the calls: two or three a day, in between the engineering work. Things were decided in them that were in no document. I meant to write each one up afterwards, and afterwards there was code to write. By Thursday, what we had agreed on Monday was a line I half remembered, and the week's calls had run into one another. Nobody can keep that up by hand and still do the job.
So the idea was small: record the call, and sort it onto the same shelves as everything else.
The idea in one picture: the call's note goes onto the same shelf as the pages about the same project, client or person.
My own thing, at first
I made it for myself. I knew the cloud notetakers and the bots that join a meeting in your place, and dictation apps that turn speech into text. Neither was what I was after: I wanted the recording on my own Mac, with nobody else on the call, and notes organized well enough to work from. It grew from there, and at some point I decided to write it properly, for one platform.
The Mac gives an app a great deal, if the app is built for it. It can take the sound of a call app on one track and the microphone on another, with no bot in the call. It knows the calendar, so the people on the call have names. It has Reminders, so an action item lands where you will see it. It has a Neural Engine for transcription and a GPU for the language models. Patum uses what macOS offers, and runs nowhere else.
Local models
The longest work went into the models. You can't run anything huge on a laptop, and a model that fits in a few gigabytes doesn't write like the big cloud ones. It can still do far more than it seems, if you stop asking it to be right in one pass.
So Patum doesn't take the model's word for a note. It first reads the transcript into a list of facts, each with the second and the speaker it came from. It writes the note with that list beside it. Then code checks the result: a number that isn't in the transcript is taken out, a date the model guessed becomes a blank, a decision from the list that the note missed is offered back, and every line is read against the transcript lines it came from, so the right person is named. The facts that survive are underlined in the note, and each one opens the moment it was said. It took many attempts and several approaches to get there, and it still does with every change of model.
The team agreed to ship the queue migration behind a flag on Nov 3 and to keep the v1 API until January. Who runs the Q1 load test is still open.
Click an underlined fact to see who said it and when. The checks are code, not the model's word.
Who said what
The other long piece was telling the people on a call apart. The models for this are good with a few voices and stop at a fixed number: four for one I tried, eight for the one Patum uses. In a town hall, the ninth person has nowhere to go.
Patum runs the model on sixty-second windows of the call instead, and a second model links the voices across windows by how they sound. Inside one window there are at most eight voices. Across the call there is no limit. On my own meetings, and on town halls with fifteen people talking, that is what held. On a call, your microphone is a track of its own, so your words are never given to someone else.
One window holds at most eight voices. The call holds as many people as there are.
What it is now
Patum is still the tool from that new job: record the call, write a note I can check, put it on the shelf. It runs on your Mac, and your recordings and notes stay there unless you choose otherwise.
Nikita Orobenko
How Patum works shows each of these steps as it is today, and what leaves your Mac.