ENGINEERING VERSION LOG · 2026.07

From dictation to a Voice Inbox

The mobile product should not duplicate desktop dictation. One recording becomes a useful object while preserving what the user actually said.

InkTyper Engineering2026.07.31

4 objectsNote, task, idea, and meeting

Raw-firstDerived structure never replaces the transcript

50–100Representative clips in a pilot

2 deliveriesHosted API or private Docker

ScopeProduct and API experimentThe mobile direction has since narrowed

EvidenceArchitecture hypothesis plus reference clientObserved results, targets, and estimates are kept separate.

StatusNot deployed as the desktop primary pathThis log records a dated engineering state, not an evergreen promise.

01 · PRODUCT HYPOTHESIS

Mobile needs its own interaction model

Desktop InkTyper sends speech into the active text field. Mobile is better framed as a Voice Inbox: each recording becomes a note, task, idea, or meeting record and then enters the appropriate workflow.

02 · DATA CONTRACT

The raw transcript is the immutable fact layer

Classification, titles, due dates, and formatting are derived views. They may be edited or regenerated, but never silently replace the original speech. The contract returns transcript, objects, quality, and timing_ms together.

03 · REFERENCE CLIENT

Validate the full path in a real client

Authenticated cloud transcription.
A local Voice Inbox with four object types.
Editable transcript preview, copy, and system sharing.
A simple local classifier first; schema-bound API output in production.

04 · B2B PILOT

Turn internal capability into a bounded evaluation

A pilot uses representative Chinese and English clips to compare latency, transcription quality, and structured output. Delivery may be a hosted endpoint or private Docker deployment; a technical evaluation comes before a commercial commitment.

EVIDENCE · DIAGNOSIS

What the evidence establishes—and what it does not

The experiment established one durable boundary: the raw transcript is the fact layer; tasks, titles, summaries, and formatting are derived views. Mobile should not copy the desktop overlay. It needs system-input, voice-inbox, and follow-on object workflows of its own.

Reading ruleProduction observations describe the observed path. Controlled benchmarks isolate a component. Targets remain targets until a sustained distribution proves them.

OPERATIONS · OPEN WORK

Failure behavior and the next verification gate

Android can use an InputMethodService for system-wide entry. An iOS third-party keyboard cannot capture the microphone directly, so a companion app, Shortcut, or system entry point must record before the keyboard extension inserts a recent result. The B2B contract must preserve transcript, derived objects, quality, and stage timing.

Failure visibility: Every fallback needs a reason code and stage timing. A successful final transcript must not erase evidence that real-time, AI, or paste failed first.
Release evidence: Keep request ID, actual model, node, queue, upload, ASR, AI, and delivery timing together so a regression can be reconstructed end to end.