Handy sent me down a rabbit hole. I came back with a fox.

[dulevu]

Hi Handy folks, and hi to the person who built it,

I'm a radiographer, not a software engineer, and I want to say thank you.

How I got here. Late one night I was doomscrolling and landed on an Instagram creator (@itstundeanthony) presenting an AI skill. Something clicked. My first idea was a voice tool for radiologists, but that already exists and getting anything into a hospital system is a wall. So I looked at my own workday instead: my colleagues and I lose time typing short patient histories, again and again.

Somewhere on that path I found Handy, expecting to play with it for ten minutes.
What Handy gave me. The proof that fast, local, offline speech-to-text is not a lab trick: it runs on an ordinary Windows PC, offline and it just works.
I used MIT-licensed parts of the Handy ecosystem (the local transcription engine and the microphone library) as the technical starting point, kept the licence notices, and credit the project in my own documentation.
My project, Fritzo, is independent and private; it is not Handy and Handy did not endorse it.

What I made with it.
One very narrow thing: press a key, say two or three sentences in Italian, press the key again, and the text is on the clipboard for me to paste myself. No cloud, no accounts, no stored audio or text, no automatic pasting. It has a small fennec mascot whose ears light up when you speak. I tested it with invented dictations only. It is a prototype, not approved for real patient data, and it still needs a human to review every sentence.

Three things that might genuinely help you or other users (observations from my setup, I may well be holding it wrong):

• Silence can come back as confident gibberish. In a quiet room, one run returned a whole line of invented Cyrillic. A tiny "was there actually any speech?" check before the model runs, plus rejecting results in the wrong script, fixed it for me. An optional, documented voice-activity step could save others the same surprise.

• Portable builds. An engine compiled on one PC can quietly target that exact CPU, which is great locally and risky for anyone shipping an installer. I pinned a safe baseline and measured no slowdown. A line in the docs could help people who package apps.

• Predictable model lookup. A clear search order (installed resources, then next to the executable) made packaging calm and avoided hard-coded developer paths.

How I checked it, in one breath: automated tests (25 of 25), 50 simulated start/stop cycles (50 of 50), 60 rapid key presses without a crash, zero network connections in normal use, and plenty of my own dictation with the laptop microphone.

Two questions I'd love your thoughts on: how do you shrink an offline app when nearly all of its size is the speech model, and how do you update a fully offline app safely, without anything phoning home?

The full story, the bugs testing found, and what I decided not to build are written up in detail on my profile: github.com/dulevu

I'm not sure this exact idea can become anything bigger, and I say so openly. But I'm actively exploring where this way of working, small, private and verified, could matter more. If you're working on local, privacy-first AI tools, I'd really like to talk.

Thank you for building something that let a non-programmer get further in days than I expected in a year. If anything above is wrong, tell me and I'll fix it.

(P.S. The fox has ears that light up when you talk. Yes, it's necessary.)

Apparently, this is what I do now haha