We deleted our backend.

The version you'll never see

ParseMeet 0.x was a cloud app.

The first builds worked the way almost every meeting-notes product works: record on the device, upload to a backend, transcribe with a hosted speech-to-text API, summarize with a hosted large-language-model API, store the results in a database, sync them back. It worked. That was the problem.

Meeting audio is about the most sensitive data a work tool can touch — your negotiations, your team's unguarded arguments, your client's confessions about their budget. Our architecture meant every minute of that would travel to servers we operated and to AI vendors we contracted. We could write careful policies about it. We could encrypt it in transit and at rest. But "encrypted in transit" still means it leaves, and a policy is a promise, not a property.

The pivot

Then Apple shipped the missing pieces.

With Apple Intelligence, the two hard jobs in this product — speech recognition and language-model structuring — became things a Mac and an iPhone can do locally: Apple's on-device speech models for transcription, and the Foundation Models framework for turning a raw transcript into decisions, action items, and owners via guided generation. Slower than a data center? Somewhat smaller model? Yes. But the job here is extraction and structuring, not essay writing — and the output only has to be right, not poetic.

So in mid-2026, before ever shipping publicly, we deleted the backend. Not migrated. Deleted: the servers, the upload pipeline, the accounts system, the API keys — roughly 3,100 lines of networking code removed from the app. A check in our build gates still fails the build if networking symbols ever reappear. The privacy policy got shorter because the architecture did.

The second chapter

Then Apple built the only cloud we'd trust.

We wrote above that a policy is a promise, not a property. Apple's Private Cloud Compute is what it looks like when someone makes the promise a property: per Apple, its silicon servers never store your data, run software published for independent security research, and refuse connections from your device unless that exact build has been publicly logged. Not even Apple can read what they process, per that same published guarantee — and we never could, because there's still no ParseMeet server anywhere. So when a meeting needs more than the on-device model, ParseMeet uses the bigger Apple model on PCC, and the privacy story doesn't bend. The same standard applies to sync: your meetings move between your Mac and iPhone through your private iCloud database — invisible to us by design, end-to-end encrypted under Advanced Data Protection.

What that buys you

Properties, not promises.

• Offline always works. Recording, transcription, and on-device parsing run with Wi-Fi off. A connection only ever adds — it's never required.
• No subscription, because no meter. Vendor AI has a per-minute cost someone must pay forever. Apple's models don't, so ParseMeet is one payment.
• "Data Not Collected." That's the App Store privacy label ParseMeet launches with — there's nothing on our side to collect with, and the app contains zero third-party SDKs.
• Pressure-tested. Before release we ran an internal adversarial security audit over the whole codebase; the no-backend claims held. The cloud we do use is the one Apple opens to independent audit.

What it costs

The honest ledger.

This architecture has real prices. ParseMeet needs Apple Intelligence hardware — an M-series Mac or an iPhone 15 Pro or later. The on-device model is smaller than the frontier models in data centers; we'd rather give you a fast, private B+ on every meeting than a slow, metered A-. And there's no team cloud: your notes are yours, on your devices, which is the point — but it means we'll never sell you an org-wide transcript archive.

1.0 is the floor, not the ceiling. But the foundation is settled: what happens in your meetings stays yours — on your devices, in your iCloud, readable by no one else.