Case study 03 WREN
Case study 03 · 2026
Dictation that never leaves your computer. Hold a key, talk, done.
01 — The problem
The well-known dictation tools are subscriptions that send your voice to the cloud. The private, local ones are mostly Mac-only. And every one of them mangles names, jargon and spellings it has never heard.
WREN fills that gap: hold a key, talk, and the words appear at your cursor in any Windows app. No account, no server, no audio leaving the machine, and it learns your words. Pay once, not every month.
02 — The decisions, and what they cost
A local language model took 696 ms to tidy a sentence and made it worse by deleting correct punctuation. A simple rules pass took 0.063 ms and kept it. Cost: no magic “rewrite it nicely”. Gain: about 11,000× faster, it never rewords you, and it can't hallucinate.
Every speech model I tested got the same product names wrong in different ways. A bigger model doesn't fix words it has never heard. GLOSS, the vocabulary engine, is a simple table you own. Cost: you maintain a list. Gain: instant, exact, yours.
$49, verified on the device. Copy protection would need a server that phones home, which breaks the product's main promise. Cost: a key can be shared. Gain: the privacy promise stays true, with no exceptions.
Windows quietly blocks unsigned downloads. Packaging WREN for the Store means Microsoft signs it. Cost: a certification queue and two rejected rounds, both on listing text, never on the app. Gain: a real install: Start menu, its own icon, no warnings.
On launch night the website download wouldn't install on a clean PC. I stopped the campaign the same night and set a new rule: nothing ships until a stranger's install has been walked on a machine that didn't build it. Cost: launch momentum. Gain: the first impression is the real product.
03 — What broke, and how I fixed it
Distribution
Windows marks downloaded files, the mark spread to all 3,472 files in the app, and a security feature blocked it. None of this showed on the machine that built it.
Fix: launch pulled, repackaged as a Microsoft-signed Store app, and a new rule that every release is installed on a stranger's machine first.
Testing
The graphics-card speech model loads fine on a PC without the right drivers, then fails on the first real sentence. The fallback to the processor never triggered.
Fix: the fallback now reacts to a real transcription, not just to loading. Found while packaging, before any customer saw it.
Reliability
WREN starts with Windows, and its first model load could lose a race with the computer's own startup. It tried once and gave up.
Fix: retry with a growing pause that still quits instantly, plus a fault log with the full error, which then diagnosed the next bug in one line.
04 — Results
Status, October 2026
Published in the Microsoft Store in October 2026 as “WREN Local AI Dictation”, publisher Vonfal AI, free to install. The app itself passed every review round; the two rejections before launch were about listing text, never the package.
05 — What I bring from this