Privacy claims should be architectural.
"We respect your privacy" is a promise. On-device-first processing is a mechanism. Regulators and customers are learning the difference fast.
The Lab · Case Study 03
Say it. It's handled.
You'll forget this. The thought you had at 11pm, the thing you meant to do, the person you meant to call. It deserves better than your memory. Dex is a voice-first execution engine: talk for ten seconds and it becomes your tasks, your people, and your morning brief.
One rule shaped every architectural decision: your voice never leaves your phone, and it never trains anyone's model.
Private betaiOS + AndroidVoice-first
Read the essay: You'll forget this →
Visit dex.builder-singh.com ↗Landing page with a live, working voice demo. Try it.
The business case: capture beats organization
Every notes app, task manager, and journal on earth has the same business model: organize what you remembered to write down. The leak is upstream, in capture. The moment of intention has zero friction for speech and high friction for typing, so the world's most important inputs evaporate in the gap.
For consumer AI the market question is sharper. People hand their most private thoughts to a system only once. Privacy isn't a compliance feature here, it's the product. A voice assistant that might leak, train, or resurface your worst moments is a voice assistant nobody keeps.
Dex is the lab's answer: a private voice journal that writes itself. On-device speech recognition turns ten seconds of talking into a structured entry, and AI extracts tasks, people, places, and commitments automatically. Text only, never audio.
A journal that writes itself. Remembers your people. Handles the to-dos.
How it works, and how it's built
The loop is three moves. Capture: ten seconds of speech, transcribed on the device, the audio never uploads. Extraction: AI turns the transcript into tasks, memories, people, and places in seconds. Execution: everything lands where your day actually happens, including a 6:30 morning brief built from your calendar, overnight email, and open loops, sorted into what you MUST do, SHOULD do, and MAINTAIN.
It runs real households, not a demo family: shared lists, delegation, rooms. The engineering underneath is deliberately boring and heavily tested. A Flutter app on iOS and Android, a Cloudflare Workers + D1 backend, and an LLM gateway for text extraction. The extraction pipeline carries 52 golden tests that replay on every change.
The telemetry proves the privacy story: analytics record screens and events, never note content. Session replay is disabled in code, not by policy. When a privacy claim is architectural, you never have to argue it.
The architecture, briefly
- Voice capture
- On-device speech recognition first, on iOS and Android. Audio never leaves the phone; cloud transcription exists only as a budget-capped fallback
- Extraction
- An LLM gateway turns transcripts into tasks, memories, people, and places. Text only. Ten to thirty seconds from speech to structured data
- Morning brief
- 6:30 daily: calendar, overnight email, tasks and goals, synthesized into MUST / SHOULD / MAINTAIN with sources
- Backend
- Cloudflare Workers + D1, free-tier-first, designed for real households with shared lists and delegation
- Quality
- 52 of 52 pipeline golden tests passing. Release gates on every build. Telemetry carries events and screens only, never note content, never audio
What this proves for client programs
Voice is an enterprise interface too.
Field teams, clinicians, warehouse leads. The capture gap exists wherever hands are busy, and the pattern generalizes well past consumers.
LLM pipelines need golden tests.
Consumer-grade reliability came from replaying 52 golden cases on every change. The same eval discipline as FundHouse, in a different domain.
Where it stands
Dex is in private beta with a small set of real households, the toughest users I know. The landing site has a live voice demo you can try in one click. When the beta opens up, this page becomes the product page.
The bridge to your program
Building consumer AI that people actually trust?
Trust is won on what the system structurally cannot do, not on policy pages. I help teams design privacy-first AI products where the claim and the architecture are the same thing.