Started the week by deleting code instead of adding it. A couple of settings in my side projects nobody ever touched, and each one was one more thing to test and explain. Less to maintain feels better than another feature.

Looked at my GeoCam numbers. 326 installs, only 29 turned on geofencing, and most people who saw the paywall just left. So single device geofencing is free now. Let people feel it work first, then ask for money.

Gave my small on device model 5.23 hours of sleep and it read it as 5 hours 23 minutes. Deep sleep 1.17 hours became 1.17 percent. Data was right, format was the problem. Now I send plain Xh Ym text and the app does the math.

Writing phrases for the youngest players in my Wheel of Fortune game is harder than the grown up ones. Short easy words, but still need to feel fun to guess. Simple is not the same as boring, still learning that.

Most people dropped off in GeoCam before the geofence ever turned on. So I added a quick start with 4 ready made routines instead of a blank screen. Also tracking the funnel now, to see exactly where people stop instead of guessing.

Put a lifetime option on GeoCam because my running cost is close to zero and some people just refuse subscriptions. Strange feeling pricing something once and hoping it stays cheap to run. Still feels like the fair deal.

Thinking about moving my health app to Flutter so it can reach iOS too. The health data part works with any phone or watch. The real limit is hardware strong enough to run the model on device, not the data.

My daughter is 8 and got hooked on Wheel of Fortune. Every app we found was full of ads and the phrases were way too hard for her. So I built my own. Sometimes your first user is sitting at your dinner table.

My on device model runs on CPU on my Fold 6 because the GPU isnt available. Each answer takes 15 to 22 seconds. Private and free to run, but you pay with waiting. Still not sure if people will wait that long.

Health data plus a small on device model. My rule for healthcoach is the model only explains numbers the app already calculated. No diagnosing, no guessing. Boring rule, but it feels like the only honest one at this size.

The family rule in GeoCam sounds simple. Last person out arms the cameras, first person back disarms. But once hourly schedules were also running, deciding who wins when they disagree took way longer than the actual code.

Found a nasty bug in geocam. Routines are paid, so for free users the geofence never got registered with the OS. It only worked after a full app restart. Intermittent bugs like this kill trust and you never see it happen.

I was logging temporary network hiccups as errors in Crashlytics. My real crashes got buried in the noise. Cleaned it up and the reports finally mean something. If everything is an alert, nothing is.

Moving my side projects from Codex to Claude Code. Biggest change is writing small skills for stuff I repeat every week. Less re explaining, more just saying what I want. Setup takes an evening but it pays back fast.

Tested tool calling on a small on device model. Tools with no arguments work fine. Tools with arguments, it just prints the call as text and invents a confirmation. So writes go through normal UI now, the model only reads.

Small on device models are bad at tool calls with arguments. Mine printed the call as text and invented the confirmation without running anything. So writes go through normal UI now, the model only reads. Dont trust it with data that matters.

Most people who install my app never turn on the main feature. So I made a quick start with 4 ready routines, nobody starts from a blank page now. I also added tracking to see exactly where they drop before activating.

I only ask for an app review after 2 successful geofences now, not on first launch. Ask people who already saw it work, not someone who installed 30 seconds ago. Seems obvious but I did it wrong for months.

Thinking about moving my on device health coach to Flutter so it can reach iOS too. Same Gemma model file, just a different wrapper. Health data part works on any phone, the real limit is if the hardware can run the model.

Made the core feature of GeoCam free for single device users. Out of 326 installs only 29 ever turned on geofencing, most left at the paywall without trying. People need to feel the value before they pay. Added a lifetime option too.