Skip to content
quiet publish.
Back to the world

Clarify02THE LISTENING PAVILION

Arti

Every conversation leaves a trace

Hear what matters.

A place to make sense of conversations. Find the important ideas in what was said and the next steps worth carrying forward.

In the studio

Explore this place in the world

In the world

SAMPLE CONVERSATION · AN ILLUSTRATIONIllustrative sample. The text is fictional, not real user or repository data.
“Could we release on Friday?” “Only if the review is finished.” “Let’s decide tomorrow.”
Find the thread
Idea: a Friday release. Condition: the review must be complete. Decision: still open.

Overview

From side project to studio tool

Arti started as a side project to stop losing what gets said in working conversations. It had a different name then. Once it had earned a place in our own week, it became Arti, and it has stayed close to home since: a tool we use before it is a tool we sell.

Understanding, not just recall

A transcript is a record. It is not an understanding. The hard part is the second step: noticing that a condition was attached to a decision, or that a question was raised and never answered. Arti’s summary is written to catch those things, and to show its work. Every claim points back to where it was said, so you can listen and judge for yourself.

What we are still working on

Recording is capped at ninety minutes, and very long conversations are the next thing to make robust. The quieter work of the last year has gone into trust: editing, correction rules that carry across the archive, and a pipeline that fails slowly and visibly rather than losing a conversation.

PRINCIPLES AT WORK

The problem

What matters in a meeting, a call or an interview is said out loud and then lost. Notes are partial, the audio and the transcript live in different places, and the follow-ups slip.

OUR APPROACH

Keep the audio, the transcript and the understanding together, and write the conversation up the way a thoughtful colleague would: what it was really about, what was decided, what is still open, and who is doing what next.

What it does

  • Records in the browser, keeping a phone awake through a long conversation and recovering a draft if the page closes, or takes an uploaded file.
  • Transcribes with speaker labels, and lets you rename speakers and correct the text.
  • Remembers correction rules and applies them across past transcripts, so a name misheard once is fixed everywhere.
  • Writes an editorial summary: key points tagged as decisions, risks or open questions, action items with an owner, and the moments that mattered.
  • Links every point back to its moment in the audio. Select it and the recording plays from there.
  • Groups conversations into projects, with full-text search and a timeline.
  • Looks across a whole project for themes, showing an estimate of the cost before it runs.
  • Shares a conversation through a link that can expire, and supports workspaces with roles.
  • Exports to text, Markdown and PDF.
  • Arti conversation page for 'Relaunch scope: what we cut, and why'. The left column opens with a serif summary paragraph, then 'What happened' and 'What was decided' sections listing four timestamped action items with named owners. The right column shows the transcript panel with three named speakers, a sentiment bar (43% positive, 43% neutral, 13% negative), and extracted entities. An audio player floats at the bottom.
    ScreenshotFictional dataA conversation, written up: the summary, what was decided and who owns what, with timestamps back to the audio.
  • Lower on the same conversation page. The left column lists 'What stood out' bullet points, topic chips, and 'Moments worth replaying' with five timestamps from 0:14 to 32:48. The right column shows the transcript, with each turn labelled by speaker name in its own colour and a timestamp.
    ScreenshotFictional dataMoments worth replaying beside the speaker-labelled transcript.
  • Project page for 'Tidewater relaunch' showing cross-conversation analysis. A deep purple banner reads 'What's been surfacing in Tidewater relaunch.' above a short narrative. Below it are 'Themes that kept returning' with ranked bars and 'The room's temperature' with a small trend chart marked Improving.
    ScreenshotFictional dataProject analysis: the themes that keep coming back across a series of calls.
  • Arti conversation page on a phone. It shows the project and date line, the large title 'Relaunch scope: what we cut, and why', the project badge, pills for 34:12, 3 speakers, ~2 min read and completed, the tags card, and the audio player docked at the bottom.
    ScreenshotFictional dataThe same conversation on a phone.

How it’s built

  • Next.js and React
  • Tailwind and shadcn/ui
  • Supabase (Postgres, auth, storage, row-level security, scheduled jobs)
  • Speech-to-text from Deepgram Nova-3, with speaker labels, entities and sentiment
  • OpenAI gpt-5, at medium reasoning effort for each conversation, and low, medium or high for a project by its depth
  • Hosted on Vercel

DECISIONS, AND WHY

  1. Heavy work happens in a queue, not a request

    Uploads create a transcription job, which queues the analysis. A web request never waits on a model, so a slow conversation can’t hold up the interface.

  2. One job at a time, on purpose

    When overlapping background runs began to slow the database enough to sign people out, we rebuilt the processor around a lease lock: one job per run, with every timeout set below the platform’s limit. A slow transcription now waits and retries instead of disappearing.

  3. Timestamps come from the transcript, not a guess

    The model sees each line with its time and must carry that time into every point it writes. That is what makes a decision or an action item something you can click and hear.

  4. Input is electric, output is editorial

    The 2026 redesign keeps the recording screen dark and alive and sets the finished summary like considered writing on warm paper. The design lives in tokens, and its component kit is kept in sync with Claude Design using our own open-source tool.

Timeline

  1. First version in a week: recording, transcription, analysis, projects, workspaces and share links.

  2. Direct uploads to storage and an audio player tied to the transcript.

  3. People, places and sentiment in transcripts, then a full audit, fixed tier by tier.

  4. The job pipeline rebuilt for reliability, and the “Audio, considered” redesign.

Status

In the studioClarify. Understand conversations, and the people in them.

In the studio, and in progress. We use Arti ourselves; there is no public release, page or waitlist yet.

Writing

Credits

  • Product, design and engineering: one of the studio’s founders.