We speak faster than we type.

A thought arrives as speech, dies in typing, and the tools that could catch it want your voice on their servers.

The eureka moments never arrive at a desk.

Walks, drives, the minutes before sleep. That is where the good thinking happens, and that is exactly where typing fails it. A few close friends and I had settled into a habit of leaving each other voice notes, because saying an idea keeps it alive. The layer in between is where it dies: processing the thought, formatting it, typing it out like text.

When I shared the idea of building an app around the habit, the reaction I got was not excitement. It was fear. Three worries came back, in almost the same words every time.

1. “My voice could be cloned.”
2. “My information might get leaked.”
3. “Models are getting trained on all of this.”

Those three sentences became the brief. Not a transcription app with privacy settings. A private app that happens to transcribe.

Trust you can check, not trust you have to take.

The three lines at the top of the landing page are the architecture, not the marketing. Audio never leaves your phone. No account to start. Text syncs only when you sign in. Each one is a promise the code keeps.

The writing is Whisper, an open speech model packed down until the phone can run it: 16 transcription languages, Hinglish first among them. Transcription is not a service the audio travels to. It is a thing the phone does.

One rule carries the interface: voice becomes ink.

The system is called Ink & Porcelain, and it is built on provenance. On any screen, at a glance, you know what came from your voice, what you edited, and what the machine measured.

Colour follows the same discipline. One red, and it may only ever mean recording. One green, and it may only ever mean AI. No gradients anywhere. The dark theme is not an inversion but a warm graphite tuned by hand, and the palette is locked by a unit test: slip below the accessibility floors and the build fails.

Precision alone reads cold, so the system carries a warmth layer: storybook watercolor illustrations for the empty moments, a recurring cast of a man in a sage cable-knit and his pale sage cat. The discipline holds inside the artwork too. The vintage handheld microphone is the only red object ever painted.

Halfway in, I audited my own product. It did not pass.

Two months of fast building produced an app that demoed well and wobbled in the details. I wrote the audit the way I would write one for a client: a strong 0.8 of a product, most of the way there, with trust-breaking seams in capture, in what counts as a note's source, and in a paywall that promised more than it could sell. The strategy after that day was two words, trust before breadth, and the decisions that followed all subtract.

The mark, the film, the site: one product.

A product is not just its screens. The launcher mark went through three explored directions before the porcelain speech bubble shipped. The launch film runs 59 seconds, storyboarded from ten painted stills. And the landing page keeps the promise before you ever install: it makes zero third-party requests until you press play on the film, and a test fails the site if that ever changes.

Under all of it: Kotlin and Jetpack Compose, and on-device speech recognition. The engineering matters here for one reason. It is what makes audio never leaves your phone true at the code level, not just in the copy.

An early product, claimed at its real size.

Install numbers are deliberately not claimed. The claim is the judgment: what got built, and what got cut.