Writing · 7 notes

Used. Built. Learned.

Things I’ve made, tools I’ve used, and what actually worked. Where I could measure it, the numbers are in.

3 min read

Why I built my own notch app

Two years ago I got my first MacBook Pro with a notch and wanted something like this. This year I finally built it.
Writing
4 min read

My Digital Fly’s First Lesson: Lift One Leg

I built a digital fly lab on my Mac. Its first lesson tests a small learning loop: see an image, choose an action, receive feedback, and save the parameters for next time.
Experiment
6 min read

The Buddy: A Second Way to Talk to PushSense

Push-to-talk assumes you already clicked into the right box. Buddy mode drops that assumption: click a pearl, talk, and drag the words into any window. Here is what it took to build, and the three bugs synthetic clicks caught before shipping.
PushSense
3 min read

Why I Built a Local Transcription App Instead of Using the Cloud

Dictation apps wanted $100+ a year to rent me my own voice. I looked into what that money was buying, and the answer was a model my Mac could run for free.
Local AI
7 min read

The 1.4GB I Was Paying for Nothing

My dictation app held 1.9GB of memory to produce a transcript a 181MB model produced identically on one clip. A second, harder clip complicates that. Measured on real speech, not synthetic audio.
Benchmark
2 min read

Building PushSense: A 5.5 MB Local-First Dictation App

How I went from a 200 MB Electron app to a 5.5 MB native macOS app using whisper.cpp and Swift.
PushSense
4 min read

Verified Vibe Coding: Let AI Write Fast, Then Make It Prove Itself

My workflow: use Claude to write code 10× faster, then measure whether the result actually works.
Agent workflow

More from Billy