everything

A short list of some stuff I've put together for personal use using AI coding.

Writing Tutor

A writing tutor for my kids, based on the Hochman Method, which walks them through construction of sentences, paragraphs, and so on.

Jeeves bot

A bot that sends me a daily synopsis of events, workouts, and the weather. Currently the only interactive element is a flow where I can add/remove/change the workouts for today or tomorrow that are planned in my intervals.icu account.

Tulay language bridge

“Tulay” means “bridge” in Tagalog, which I'm trying to learn. The app pulls information from my LingQ account and my Anki decks to create a master list of all words I know so far and what their status is. From there, I can generate mnemonics, for example for the word alagang, which means pet, it generated this mnemonic story:

all a gang — In a sunny yard, a whole gang of puppies and kittens crowds around you, licking your hands—every single one is your pet, and you laugh, 'All a gang? Yes, all a gang of beloved pets!'

This helps anchor the word in a funny story and aids in memorization.

Then based on the mnemonic, it generates 3 sentences which use at least 80% words I already know, and formats them into cloze cards in Anki, and syncs them to my deck.

Each morning, the app sends me a notification (via web browser notifications) with a ~100 word story that contains 80-90% words I already know, uploads it to LingQ as a lesson, and gives me the link. This pushes me to learn something new each day, beyond reviewing already known words.

A few unfinished things

• A surveillance system for my brother in law, he ended up using another programmer for it so I'm letting it go, but I might pick it back up when we move to a family compound in the province soon, as we'll have a need to collate surveillance feeds and make them actionable for us and our employees.
• A learning system for my kids for math. This ended up being a lot more complicated than I originally realized, so I kind of drifted away from it. The idea was to create a tree of all the Common Core math standards, with explicit dependency links between each standard, so if a student fails on a standard in a quiz, before re-teaching the standard, the system would first quiz all standards that one depends on, making sure there is no supporting information that is missing for the student. I have observed in my own kids that often a failed learning experience is not because of the information in front of them, it's because of a hidden assumption that they know something else or have fluency in something beyond what they have currently. For now I just have to do this manually by observing their results and intervening.

How I've been creating stuff

I've always struggled in the past with finishing personal projects. Typically I get a burst of inspiration that lasts a couple days, and in that time I can do an enormous amount of thinking and implementation, but before AI, I could only get so far before I hit some very annoying roadblock. Think like, I can't get my file permissions correct on my webserver, so I can't test the code I've written, so I'm stuck upgrading random packages and unwinding years of idiosyncratic setup on my VM. Then eventually I'd just run out of steam and drift away from that project.

With AI, that no longer happens. I just tell it what to do, and it does it. Even this blog was down for over 2 years because of simple issues that I just didn't want to deal with. I asked Cursor to fix it and it was done in 5 minutes, and I updated the WriteFreely instance to boot. In the past even that last step would have required hours of reading 2 years of release notes watching out for gotchas (Cursor time, 30 seconds), running migrations, probably getting stuck on some stupid stuff.

While writing this post, I added image hosting to my local WriteFreely instance, even though that feature has been unfinished in the official project for years.

So now, with AI, I can generate such a huge amount of code, I can actually complete things before my motivation shifts. It's a huge difference for me.

For models and tools, I use Gemini to plan stuff, and it's my daily driver “chat” tool usually as well. Then I take whatever that generated and add it to a new repo in Cursor, which I use because last October I signed up for the yearly plan that had unlimited “Auto” model mode, which is an incredible deal. I can do billions of tokens a month with no extra charge.

I have some stuff running in val.town but I plan to port it out to my personal Linode server (which I plan to move to a cheaper VPS eventually, maybe OVH). Some things I run on my home desktop and then tunnel out to the web via my Linode, like Tulay.

When I embed a model into an app, like the mnemonic generation in Tulay, I have been using Deepseek. The deepseek-v4-flash model is almost too cheap to meter at this point. I've used it for weeks and racked up less than $1 of charges so far. It's plenty smart enough for most things. It may even be good enough for coding, but I have that free Cursor usage so I just use that for now. I'll re-assess in October when my plan expires, I'd rather not be using an XAI product in the long run.