You'll forget this

It is 11pm. You are somewhere between the kitchen and bed when the thought arrives: the thing you promised your brother, the call you owe the bank, the idea for Monday. You think, clearly and sincerely, I will remember this. You will not. Nobody does. The thought dissolves overnight and resurfaces three weeks late, usually as an apology.

Every notes app on earth is built for the survivors, the thoughts that made it onto a screen. The leak is upstream of all of them, in the moment of intention itself. Speech has zero friction at that moment. Typing has enough friction to lose most of what matters. So the world's important inputs evaporate in a gap that every capture tool pretends doesn't exist.

What ten seconds buys you

Dex is the lab's answer. Talk for ten seconds and the app does the rest: transcription on the device, then extraction into tasks, people, places, and commitments. Everything lands where your day actually happens, including a 6:30 morning brief that sorts your calendar, overnight email, and open loops into MUST, SHOULD, and MAINTAIN. It runs real households too, with shared lists and delegation, because the 11pm thought is rarely just yours.

The decision that shaped everything

Here is the product truth nobody in consumer AI gets to skip: a person hands their most private thoughts to a system exactly once. The first betrayal is the last user. That makes privacy a product decision, not a compliance page, and it forced the architecture before any feature was chosen.

Speech recognition runs on the device, on both iOS and Android, so the audio never leaves the phone. Cloud transcription exists only as a budget-capped fallback. Extraction works on text, never audio. Analytics carry screens and events, never note content, and session replay is disabled in code rather than by policy. When someone asks how their voice is used, the honest answer is a mechanism, not a promise.

Built the way I tell clients to build

The engineering is deliberately boring and heavily tested: a Flutter app on iOS and Android, a Cloudflare Workers and D1 backend, extraction through a model gateway, and 52 golden tests that replay the extraction pipeline on every change. The system was architected under my direction and built by AI agents, which is the same operating model I set up for enterprise teams. The interesting part of consumer AI is not the model. It is the discipline around it.

The test I apply to any privacy claim now is boring and mechanical: trace where the bytes go, in the code, on the worst day, when a shortcut would be convenient. Promises live in documents and documents have version history; mechanisms live in the build, and the build either uploads the audio or it doesn't. That is also why the golden tests matter as much as the architecture. Every change replays the extraction pipeline against known inputs, so the property being advertised is regression-tested like any other feature. Privacy that can break silently in a refactor was never privacy. It was a mood.

The trust lesson, stated plainly

Users cannot verify promises, but they can verify mechanisms, and regulators are learning to tell the difference faster than marketing departments would like. If you are building consumer AI, find the sentence beginning with we promise and replace it with something the system structurally cannot do. That sentence is your moat.

Dex is in private beta with a small set of real households right now, and the landing site has a working voice demo you can try in one click. The full case study covers the architecture and the privacy engineering.

Go deeper: Dex case study