The work after the meeting

The follow-up to a meeting often takes longer than the meeting. You scrub through the recording to find what was decided, open the proposal someone mentioned, check a number, and write it up. Each piece lives in a different app. Most AI tools that could help ask you to upload all of it first.

For plenty of material, that's a fine trade. For client calls, interviews, unreleased work, or anything covered by a confidentiality agreement, many people can't make it, and others would rather not.

Local has to include the model

"Local" is often used for apps that store files on your device but send their contents to a remote model for processing. We wanted the model to run on the same machine as the data.

Lumen uses Whisper large-v3-turbo for transcription, Gemma 4 E2B to write meeting notes, and EmbeddingGemma 2 for search. All three run on the device. The models download once from Hugging Face during setup; after that, the app and its AI processes have no network access, and we've tested transcription, notes, and search with networking blocked. If a model is missing, the feature that needs it is unavailable and Lumen tells you. It doesn't quietly send the job somewhere else.

What we built first

We started with two jobs that most knowledge work depends on: understanding a conversation and finding something in your files.

Meetings. Record your microphone and, optionally, call audio, or import a recording you already have. During a recording, the transcript fills in as each chunk of audio is processed. When you stop, Lumen writes a summary, decisions, next steps, and open questions. You can rename speakers, edit next steps, search past meetings by what was said, and export everything to Markdown.

Search. Choose the folders Lumen can read. Exact names, numbers, and quoted phrases match directly, and PDFs open at the matching page. With the optional search package, Lumen also reads scanned pages and images, finds passages by meaning, locates moments in audio and video, and can index code. It never changes your files.

Every note links to its source

A model can mishear a name or assign an action to the wrong person. We don't think the answer is to hide that. Every summary point and decision in Lumen links to the part of the transcript it came from, and nothing leaves the app until you export it.

A source link doesn't make a note correct. It makes it quick to check, which is what you need before you send a recap to someone else.

Where it stands

Lumen is an early preview that runs on macOS today, with Windows and Linux versions planned. We've developed and tested it on one M1 Pro with 16 GB of memory. Transcription, generated notes, and search have passed small evaluations with synthetic inputs, and we've processed one saved recording end to end.

That shows the pieces work together. It doesn't tell us how accurate they are across languages, accents, and long meetings, or how they behave on other Macs. That's the next round of testing, along with live call capture across devices. There's no installer yet, and the source repository isn't public.

What's next

The next step is the work that comes after the notes: drafts that combine a conversation with the documents it refers to, still processed on your own computer and still linked to their sources.

If there's a task you'd like Lumen to handle, request early access and tell us about it. A short description is enough.