10
votes
What programming/technical projects have you been working on?
This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?
I have been fine tuning some LLMs, trained my robot arm to play chess, and I have been developing on my space simulator like a madman. I've shipped so much in the last few weeks. I have applied for some funding, but if that doesn't work, I'd love to figure out how to monetize it somehow, so I've just been working on it a lot, also making some reels again, so I can share info about it. So the usernumbers have been looking quite good recently!
Just don't pull a My Friend Aaron!
I'd say you could identify small local businesses that are too techno-sane and help them out net-wise. Maybe they need LLMs, maybe not--but building up a small fiefdom of B2B relationships can be a win-win situation. Just need to figure out how to market your services and package them into something where people can easily sign up with low commitment.
Yes! I’ve been thinking about that a lot. I just worry about oversaturation but then again, I’ve literally been working in tech for 10 years and especially doing AI/ML engineering for 5+ years.
I just used to like to talk myself out of things. I’ve gotten much better at it. It I think about what I hear from people, and seeing how Luna has developed - it literally almost ended up in my pile of cool stuff I built and that stays on my PC, but I just decided to share it and e see what people think of it, and it’s literally giving 50k people a month access to a fun educational game.
I drifted off, but my main point is, that I am a big proponent of local AI. I don’t think either side is right (AI isn’t useless and AI isn’t going to be completely eradicating all labor).
But the local model by now are so powerful, that with good engineering around it, you can have really cool upgrades to your workflow (that don’t involve replacing your labor force)
Reading your article now!
Reading that sentence gave me a 'wow, I really am living in the future!' moment!
Thanks for sharing, your life sounds awesome (and I hope it is).
Thank you! At some point it becomes so normal, if you take a step back and see what these agentic systems can do nowadays it is kind of magic compared to what we’d have imagined a few years ago.
Life is a rollercoaster I guess haha.
I especially like building systems that help me organize since my adhd sometimes paralyzes me and keeps me from starting a task, but once I’m into it, I can be working obsessively for a while.
I’ve now been developing both an AI integration for Luna and I’m looking into building it fully accessible so that you can use it even as a blind person. I don’t know where it will go, but I’ve had some ideas.
That’s another thing why these agents are amazing. Instead of making everyone click around in menus, you could allow someone to interact with your tool just by voice
I've finally pulled Ryubing (the successor fork of Ryujinx) out of my backlog and am trying to tackle performance issues.
It's mostly slow, learning C#, debugging tools for C#, and getting familiar with the code base.
I did find an issue where on the rendering thread half the time was spent creating/destroying vulkan fences. So that was neat. The fix was backing their fence abstraction with a fence pool and required almost zero knowledge to implement (I had to figure out how to reset the fence, but that's just an api call, the rest is just wiring in a thread safe pool pattern which touched all of three files).
It didn't double the frame rate as the bottleneck shifted, but it did increase framerate in slow scenes as much as 50%.
The two main issues now are the GPU thread (which translates/processes the native gpfifo buffer) and the render thread. Both of which require a much deeper understanding of the GPU stack/vulkan before I can make any traction.
So I asked my 8 year old daughter for a game idea, and she said juice factory. My input beyond that was to basically replicate cookie clicker & some other minor adjustments.
Had Claude whip this up.
https://juice-factory.qot.nz
After seeing this game and being excited a bit, she wrote a piece of paper (https://pic.t0.vc/DHOT.jpg) detailing another game to make. Basically trick or treating along a street.
Again, minor steering from me, and holy shit Claude smashed it.
A multiplayer co-op trick or treat game to collect sweets to exchange for more outfits.
https://candy-street.qot.nzhttps://halloween.qot.nz
Hopefully this stays within the cloudflare free tier. It uses durable objects for some persistence.
I've never played a cookie clicker game, but the orange juice game was sweet (if you will). Thanks for making it!
So I created an AI Agent to help my very dyslexic learning disabled father manage his older life. It's self hosted on my own server and it has full access to his phone and TV via wireless adb and MCP servers to help him troubleshoot his technology. I created a knowledgebase on dyslexia, both clinical medical knowledge and pragmatic practical self-help knowledge. So it uses this knowledgebase quite a bit to ground a lot of it's interactions with him with an understanding of Dyslexia.
Oh, and a database of tens of thousands of dirty jokes ingested from multiple dirty joke books.
I know the dangers of giving an AI unrestricted ADB access to control an Android phone and Android TV is, but frankly it's self-hosted, secure, it's got safeguards, and if it goes tit's up and blows up in my face then that's a lesson I consent to learning myself. But I'm pretty confident this particular usecase is low stakes the way I have it set up.
I also hooked him up with a Home Assistant server on a Raspberry Pi with a few rooted Amazon Echo Dots through EchoMuse so he can talk to the AI throughout his trailer just talking to the ether. Got him a few cheap wifi bulbs which I ported to be used offline through HomeAssistant.
It's already developed tools for him, now my dad can just take pictures of documents and it'll access the photo read it allowed to him. It's integrated with his calendar so whenever he gets a confirmation text or email for a doctor's appointment it automatically adds it to his calendar. Also looks up when bills are due through his email and helps him sort through his email to filter out the spam from important emails.
It's helped him cancel multiple subscriptions he'd somehow subscribed to over the years, and it installed anti-malware tools on his phone and checks regularly for malware on his phone.
And because it has built in memory, it remembers everything he tells it. So he can just talk to it like normal and it will remember things perfectly.
He's had a small little fishing boat in his garage that needed permits that's been sitting there for 3 years, and with the help of this AI he's been able to get it all properly permitted and is taking it out in a couple weeks.
He literally called me crying the other day thanking me for making it for him. Said it's fixed his phone with things that have been broken for a while, and has helped him form shopping lists, and he's started cooking things again instead of eating out because the AI will look up recipes and read it out to him and allow him to ask questions. He says it has infinite patience with his learning disability and he doesn't feel embarrassed or like a burden with it. He's been showing it to all of his friends at work who've all been amazed because it's so much different than LLMs like ChatGPT.
He had a friend over and was showing it off to him, but the AI couldn't distinguish between voices so I updated it with the ability to save voice profiles on people and remember them by name with permission and access levels, now my dad's been trying to invite everyone over to show off his trailer's AI butler.
It also helps me keep tabs on him, he's an ex-addict who's as close to sober as he'll ever get, so it lets me keep an eye on him. I built the AI with safeguards to not get him into any trouble and talk him down off any ledges if it needs to. I also monitor for any hallucinations or personality drift regularly.
My dad's a tough nut to crack, likes the novelty of technology but seriously struggles to use it. It's a relief that it's helped him this much and he's already doing a lot better emotionally recently.
I've been building a gym app. I hated Hevy and I refuse to pay a subscription just to record my workout sessions. I had been using notes/google sheets to record my workouts and while functional, is a bit of a pain. I discovered my brother had been doing the same thing for similar reasons. In his case, he had a good iOS app that then introduced monthly fees.
I started the app last year last year but dropped off after I had set up the plumbing and placeholders. I only just picked it up again recently and it's in a pretty stable state. I've been using Claude for the planning and development pairing with Gemini for plan and code reviews. That said, if I want to publish it to the app store I'll need to go through closed testing with a dozen or so users for Google to approve it. If I do get to that stage, I'll need to look at a way to monetise it as backend hosting isn't exactly free. I won't do subscriptions and I won't block features behind a pay well. I find that pretty scummy.
I initially started the project in angular as a way to upskill myself for work as we were shifting technologies but I don't really want to code at home in my free time. It's still an angular app but with a capacitor wrapper for a native Android build. Backend hosting is Firebase and I'm using CloudFlare for my DNS. Code is privately hosted on GitHub and I have automated pushes for the web version and manually build and release for the android app (I deployed this to play console for internal testing).
I do plan to attempt an iOS conversion but I will need to pick up a MacBook for that. I just integrated google health connect and plan on doing a wear os app soon.
My goal here is to build an app that suits my needs, is easy and convenient to use. When I used Hevy, I felt like I was fighting the UI as it did not seem simple to do the things I wanted, plus the wear os app sucked and at the time, there was no google health/fit connection. I also wanted to track workouts against my brother so I included a friend system where you can track against different exercises and view how you are progressing vs your mate.
The theme I pick for the app is Sisyphus/Greek, though I haven't settled on a name yet.
It's election year in NZ and with a little poll obsession I wanted to see what permutations of party coalition agreements would be possible.
So a small script to take some forecasted seat data from polling to work out which combinations of parties could achieve a majority.
Each party is given a letter, n l g a z m o, and a coalition is just an anagram of those letters, eg lmao for Labour+Maori+Act+Opportunity.
The permutations string was AI given because meh. The rest of this prototype.js file was my doing.
https://github.com/jason-s13r/1j.nz/blob/main/src/lab/coalition-calculator/prototype.js
That said, I did then take the prototype.js and basically hand it over to Claude to build a UI for display use on my website so anyone can play with it. And added more polls and historic election results, data from Wikipedia because ez aggregation source.
https://jason.schwarzenberger.co.nz/lab/tools/coalition-calculator/
I've been "developing" a touch-first environment for my Surface Pro 3 tablet. It's long in the tooth and low on resources, so I've been trying out some lightweight compositors like MangoWM. Compared to a full fledged DE like KDE Plasma, it runs really smoothly. I have it in a usable enough state that I feel comfortable taking it out of the house without a keyboard attached so now it's mostly just tweaking things.
I'm relatively new to tiling WMs, so it's been a bit of a learning experience for me. My two biggest issues are that tapping in an inactive tile doesn't activate it (it passes the touch event to the application instead - a physical mouse properly activates the tile when clicked in) and WiFi is flaky when resuming from sleep. Also my AuDHD is preventing me from properly sitting down and writing out a feature list. I guess I've just made up my mind that I'll add things as the need arises...
It's been a fun project though and I like that I have a rather nice tablet with a great screen, decent battery life, and pressure sensitive pen support for around 100$. It'll never be as nice or optimized as an iPad, but I can also do so much on there that I can't in the garden of Apple.
As a musician, I've always loved radio. I spent the weekend finally deploying icecast/liquidsoap on my VPS as a proof-of-concept. It's just looping a folder-as-playlist right now, but the radio is a small part of a larger project of mine -- a terminal program that will interrupt that playlist and stream live with one command.
Setup reverse proxy with nginx, fixed cross-origin problems and learned a lot -- it took a long time but I didn't want to use AI for this. More than I can say for the other thing I worked on today -- had ChatGPT coach me through rewriting a PowerShell script I use for triggering powerpoint, pdf and video files for events in C and compiling in Windows. I had high hopes of learning C this year, but it's feeling increasingly out of reach.
The program works great, and debugging it with the LLM was definitely educational.
The big music project, integrating radio, archive, live show metadata, chat, and all kinds of other stuff will be written in Elixir and I'm currently taking a course to try to really understand how it works. Been a lot of fun so far.
Weirdly I've been working on a coding harness. I find when I'm working in Claude Code, I have a sense of anxiety when I step away. Did I forget a detail? Did something important escape in the chat? Did we stick to the design? Did we do all the todo items?
Instead, I'm orienting around the planning doc. Each project is a set of tabs and each tab is centered around a planning doc which you collaboratively edit with the LLM. When you're done you put it in "work" mode which causes the plan to do be performed. It's a subtle mind shift but so far has been pleasant.
We have a Nest thermostat. I wanted to see a graph of what the thermostat was set to compared to the actual temperature. Turns out the way to get this is to sign up as a Nest developer (paying them $5), create a Google Cloud project, and add myself as a beta tester, so I can grant myself access to an API via OAuth. So now I have a personal web app running in a Linux VM on exe.dev that polls it every 10 minutes and saves the readings to a sqlite database. There’s a private website where I can see the graph. exe.dev handles authentication and I shared a link with my wife.
It’s ridiculously over-complicated for what it is, and yet was fairly straightforward to do. I used AI to walk me through signing up. Then I prompted it for a list of the commits needed to build the website, and told it to go ahead and implement them. It was working in an evening, after a lot of tweaking because AI is terrible at web design but will do what you say. I asked it to use Bun to build the website this time, but it could have been anything.
I can visit the website from any device and when I want to make the graph a little fancier, I can ask the AI to change it. There are no ads and the UI will never change unless I tell the AI to change it.
Maybe someday people won’t have to go through all this nonsense to install a personal app that they can use from anywhere and have full control over?
Meanwhile, the halloween costume is going poorly. Turns out it’s not so easy to design and 3D print a helmet.