#trial and error
7 posts- 3 min
Listening twice — when English broke the transcript
The Japanese play-by-play came through fine. But the moment an English comment began, the transcript collapsed into "Wel actaltlysputure…" — a string that looks like English and isn't. A transcription model can only listen in one language at a time. So I decided to have two of them listen to the same audio.
- 3 min
"OK" doesn't mean it opened — chasing a PDF that never appeared
Open a Mac file from the iPhone. Run it, and the log says "ok." But on the Mac's screen, nothing happens. The hunt for the culprit ended somewhere unexpected — an app that had been left running for fourteen days. And the real lesson wasn't about the culprit. It was about the "ok."
- 3 min
Speak here, type there — when I stopped asking the Mac
One button at my hand, and the Mac's dictation wakes up — that was the whole plan. But no matter how I asked, the Mac wouldn't listen. After enough stubbornness, I stopped asking. Don't make the Mac listen; let the iPhone do it. At the end of the detour sits my favorite feature in the app.
- 3 min
The meeting that never happened — teaching the summary not to invent
A recording of nothing but the word "Yes." came back with a full meeting summary attached — decisions, to-dos, the lot. Summarize a 14-minute conversation and you'd get "28 decisions." The real number was zero. The AI was saying more than it had heard, again. What I fixed wasn't the model. It was how I was asking.
- 3 min
The name was a sticky note — rebuilding speaker identification
Enroll a voice, and that person's lines get their name — that was the feature, supposedly. Except recordings of people I had never enrolled were coming back labeled with enrolled names. Digging in, I found the model had no mechanism for matching voices at all. The name was a sticky note on a seat. The most surprising story of this series.
- 3 min
Harder to read than a card — a table splits into two lines
A receipt's line items run name on the left, amount on the right. To the human eye, it's clearly one row. But text recognition sometimes returned that name and amount as separate fragments. What looked like one row split into two — and that "table" was the part of this app I struggled with most.
- 4 min
The background that doubled — cutting out a face with only Apple's public APIs
Turning a face into a drawing — that part worked. The problem was putting the drawing back onto the photo: only the face swapped, the background and clothes left as they were. It didn't come easily. The background doubled, I hit two dead ends, and still cut my way through with only Apple's public APIs — the record of one day.