MindLake is a Mac app for people whose work happens in meetings. It records, transcribes and diarises what was said, attributes the words to the people in the room, and writes the result into a note: summary, topics, decisions, action items and open questions. Every note is a plain Markdown file in a folder you choose, so the app can be thrown away and the work still reads.
The whole thing runs on the machine. Transcription is NVIDIA's Parakeet model on the Apple Neural Engine, with word-level timestamps and per-word confidence, and the summaries come from Apple's on-device Foundation Models. Audio is never uploaded, there is no account to sign up for, and a retention policy clears recordings on whatever schedule you set. A cloud summariser is there if you want it, but only if you opt in.
It learns as it goes. Speakers get named once by linking them to your Apple Contacts and are recognised by voice after that, and a lexicon of your own vocabulary, matched phonetically and by edit distance, quietly fixes the domain terms that transcription always mangles. Around that sits a Markdown editor with inline autocomplete and a heading rail, a force-directed graph linking notes by shared tags, people and folders, and a share extension for pulling in Voice Memos and files from Finder.
The app is SwiftUI over a Swift package holding all the logic, so the iPhone companion reuses the core rather than forking it. The site at mindlake.app is separate: Next.js with the App Router and Tailwind, on Vercel, with the waitlist going into Supabase. The app itself is free, and the plan is to charge only for publishing notes to the web.
- Role: Design and build
- App: Swift and SwiftUI, macOS 15+
- On device: Parakeet transcription, Apple Foundation Models
- Site: Next.js, Tailwind, Supabase, Vercel
- Live: mindlake.app





