Octo · a conversation lab · crowdfunding after the beta
Octo records your conversations, works out who said what, and reviews each one against principles you write yourself. Then it suggests one or two small things to do differently next time.
It runs on your own Mac. The audio never leaves it.
What it does 10 m
A small recorder on the table, or the voice memos on your phone. The recorder shows when it's recording.
On your Mac, Octo writes down every word and works out who said what.
How the others seemed, what went unsaid, how you did against your principles, and one or two next steps.
Confirm who's who, mark what was new, tick what you did. Over weeks, a coach names the one habit most worth working on.
How it works 30 m
Like an octopus's arms, each of Octo's eight steps runs on its own and does one job, on your Mac. You are the head.
A made-up 1:1 between you and Meera, a team lead, about a launch date.
The recorder sits in the room and records onto its memory card in one-minute pieces. When it's on your home Wi-Fi, your Mac collects the pieces every five minutes. Voice memos from your phone can go in too.
A Seeed XIAO ESP32-S3 Sense board with its own microphone, recording at 16 kHz. Each piece is encrypted on the card so only your Mac can read it. Every request between the Mac and the recorder is signed, and a piece is deleted from the card only after its SHA-256 checksum matches on the Mac. Recordings are kept as FLAC, which is lossless and about a third the size of WAV.
LimitAcross a room, voices arrive 15 to 35 dB quieter than on a phone held close. People still hear every word; the models start missing them sooner.
A day of recording is mostly silence. Octo skips it and keeps the stretches where people talk. Three minutes of quiet ends a conversation.
Silero voice detection runs on a louder copy, after an 80 Hz high-pass filter. The board's microphone drifts below 5 Hz, and until that was filtered out the detector found no speech at all. A conversation needs at least 2 seconds of speech.
LimitTwo meetings back to back can come out as one. You can split them where the people change.
Every word, with the moment it was said. First Octo works out the language, then it writes the words.
Whisper large-v3 votes on the language over the five most speech-heavy stretches, choosing among English, Hindi, Kannada and Gujarati. Qwen3-ASR-1.7B writes English; it beat Whisper 7 to 3 in a blind test on real conversations. For Hindi and mixed conversations, both models write a draft and the review's model merges the two from their text. That won 6 of 9 in a blind test, with no invented words.
LimitKannada and Gujarati are paused. Names of people and places are the weakest spot for every model tried.
Octo splits the sound into voices, A and B here, and marks when each one is talking. It doesn't know their names yet.
pyannote 3.1 on the Mac's GPU, with a clustering threshold of 0.705. Two newer models were tried on real recordings and each lost a person, putting two people on one voice, so 3.1 stayed.
LimitSimilar voices can blend. One voice split in two takes one click to merge; two people on one voice take longer to fix.
Each word goes to the voice that was talking at that moment, and the words become lines in a transcript.
Words are placed by how much they overlap each speaker's turns, and short gaps are filled by comparing voices. A line whose speaker is uncertain gets a ⚠, so nobody builds a conclusion on it unknowingly.
LimitIn fast back-and-forth, a word can land on the previous speaker's line.
Octo compares each voice with the people you've named before and suggests a name for each. It accepts your own voice on its own; other names wait for your yes.
Voiceprints come from WeSpeaker ResNet293. A name is suggested at a similarity of 0.60 or more, at least 0.15 ahead of the next person. The owner's own voice is accepted on its own at 0.70 with a 0.25 lead, and adds no voiceprint until they confirm it. Prints from under 8 seconds of speech are ignored, and each person keeps at most 12.
LimitSomeone else gets a voiceprint only once you record that they agreed. Without it, you type their name by hand.
Once A is you, A turns blue: you are always the accent colour in the app. Try the buttons.
Talk share, turns, questions, interruptions and filler words, for everyone in the conversation.
Plain arithmetic, no AI. Questions come from punctuation and interruptions from overlapping speech, so both are marked approximate. Filler words are counted only where the transcription kept them.
LimitNumbers are never coloured good or bad. The review uses them where they say something.
A model you choose reads the whole conversation and writes a review: how the others seemed, what went unsaid, how you did against each principle, and what to do next.
Claude Opus 5.5 on your own Anthropic key, any OpenAI-compatible service on yours, or a local model through Ollama. Only the transcript text, names and numbers are sent, and only to the service you chose. Every quote is checked in code against the line it cites, and an item that loses its evidence is dropped. A reading of someone else needs two of their own lines from different moments.
LimitTone is inferred from words, pauses and turn-taking, and the review says so when it does.
Her 03:12 question never got an answer. The date was held at 17:40 without coming back to QA.
Ask Meera what QA needs to hold the date, and hear the whole answer first.
Then the head. You answer any suggested names (the review waits for that), mark what was new and what you did. Once you've answered a review it stays as it is, even if the transcript changes later, unless you ask for a redo.
What you get 60 m
An example: the conversation and every line in it are made up.
6 to rateKeys work on the highlighted card:R P W right, partly, wrongN M new, knew it
Summary
A 24-minute 1:1 about the launch date. Meera raised QA capacity twice; the date was held without that being answered. You took her frustration about hearing things last well. You did most of the talking (68%).
Between the lines
You held the date at 17:40 without coming back to QA's two missing people. Her “Sure. Makes sense.” came after a four-second pause, so it may be agreement on the surface only. That reading is inferred from the pause and the short reply.
Your conduct
When Meera said the team always finds out last, you answered “That's fair. Tell me more.” and let her finish. That's valuing the colleague who gives you honest feedback.
What to do
Open your next 1:1 with her question from 03:12, and hear the whole answer before you suggest anything.
Why: she raised it twice and it was never answered.
Reading Meera
Twice she offered a way into the QA problem, and both times the conversation moved on. A direct question back may get you the whole concern.
At 21:10 you leaned on your track record, and the QA question closed there. Whether ego was part of it only you can tell; the transcript can't show what you were thinking. The effect was that her concern went unanswered.
No clear moment for principles 1 to 10 or 13 in this conversation, so nothing is said about them.
She offered one at 11:05. Ask for it this week and read it before the steering meeting.
Why: it turns her concern into something you can act on.
Limits. Lines from 08:40 to 09:10 have an unsure speaker (⚠). Tone is inferred from the words, pauses and turn-taking. Nothing here can say what either of you was thinking.
Every quote is checked in code against the line it cites. An item that loses its evidence is dropped before you see it.
Your principles 90 m
Most coaching measures you against getting ahead. Yours may ask for something else, and when the two disagree, Octo goes with yours. Advice you never want, such as to claim the credit or win the room, is never given, not even softened.
Integrity is what makes people trust you.
Stay calm; hostility is not the answer.
Watch your own four faults: anger, jealousy, deceit and ego.
You write yours when you set up Octo, from whatever you measure yourself by. Octo never writes them, edits them or suggests any, and neither do we.
Octo began in September 2026 as a tool for reviewing one's own conversations, with three goals.
Feedback on how you actually show up in conversations.
From people who have already solved the problems you are facing now.
Not built yetIn life and career, with conduct measured against your own principles rather than generic management advice.
The standard
Every conversation is judged against your principles first. Career feedback still comes, but only when it fits them. The brief every review follows: review the conversation I had with others and give me feedback the way a perceptive colleague would
.
Advice a typical coach might give after that conversation, checked against that rule:
keep a suggestion only if it passes every conduct principle. If a useful career point would conflict, leave it out rather than soften it.
Your voice 120 m
No accounts and no servers of ours: nothing you record, write or set ever reaches us.
Shows when it's recording. Sends pieces only over home Wi-Fi, signed, checked before they leave the card.
Kept as FLAC, and deleted 60 days after it's added.
Only for conversations you're in, once the speakers are settled. On your own key, or a local model so nothing leaves at all.
Title, details, due date and a link back.
The recorder's screen shows the octopus listening, with REC and the time, and its light pulses every five seconds while it records. There's no hidden mode.
Audio, voiceprints, transcription and telling voices apart all happen on the owner's Mac. No cloud service hears anything, and that isn't a setting.
The transcript text, names and numbers, on the owner's own key; or nothing at all with a local model. Only for conversations the owner is in. One they aren't in gets no review, just numbers.
Someone else's voiceprint is kept only once the owner records that they agreed: when, and how. Without it, names are typed by hand and no voiceprint is made.
Transcripts and reviews stay, on the Mac. Deletes reach its backups within 14 days.
It describes what others did and seemed to want, in their own words, so the owner can understand them and respond well. Never a verdict on anyone's character or worth.
Anyone recorded can ask the owner for any of these, at any time. Each is a button in the app, and each asks before it acts.
The recorder 150 m
It sits on the table in plain view and records until you stop it. This one works: try the controls on its side.
Starting
Tap record to wake the screen; hold it 2 seconds to start or stop.
A drawing of the case design, not a photo. M1 is the Seeed board and EB the expansion board, which carries the screen. M1 stands on EB, which makes one end taller. The lower level's top is dark glass, with the screen and the recording light under it. The working recorder runs today as bare boards on a power bank; the case is designed and not yet printed.
The screen on this page uses the recorder's own drawing code, ported line by line and checked against the original: 1,080 frames, not one pixel different. On a 128 × 64 screen the octopus makes do with six arms.
The plan 170 m
Crowdfunding only once it has shown it works, against a bar written down before the results.
Does it tell us anything we didn't already know? 10 reviews, 1 in 3 with something new, 2 acted on.
On their own Macs, with their own principles, for four weeks.
For the recorder, assembled in its case.
Where stage one stands
The beta begins, and the case gets printed and fitted.
We stop building, and there is no campaign.
The beta passes if people install it without help, at least 1 in 3 of their reviews has something new to them, and they're still using it in week 4. Before the campaign opens: a printed case that fits, certification for the boards, and a production plan for a first batch, shown on a working prototype.
Follow along 190 m
Two kinds of news, rarely: a place in the beta if you have an Apple Silicon Mac, and the day the campaign opens.
This copy of the page can't keep a list. Write to support@octolab.online and we'll add you.