site banner

Tinker Tuesday for June 10, 2025

This thread is for anyone working on personal projects to share their progress, and hold themselves somewhat accountable to a group of peers.

Post your project, your progress from last week, and what you hope to accomplish this week.

If you want to be pinged with a reminder asking about your project, let me know, and I'll harass you each week until you cancel the service

2
Jump in the discussion.

No email address required.

I've been tinkering with LLMs recently.

I work as an engineer and I have been trying to build a second-brain repository of information in a database that I can use as context for an LLM to query. Something like Perplexity.ai, but on a local machine because I'd be working with company data. In the last week or so I've uploaded PDFs of whitepapers, books, and industry standards. So far, I've found that the LLMs are miles ahead of plain text search. The platform provides citations so the underlying references can be found quickly.

I am using a standard install of Open WebUI and ollama on a Linux machine. I've tried various smaller models (Deepseek-r1, Phi-3, Phi-4) and have been generally successful, but for larger models find that I just don't have enough computing power. I am comfortable installing and setting up software in a terminal, but I have no formal coding/software development background. So if I can do this, you can too.

Next week I plan to upload several years' worth of e-mails into the database, and see if I can run queries against it.

I wonder how much information can I upload into a database before I start running up against constraints?

If you're using any sort of real database (sqlite / mysql / postgres / mongo / etc) to store the documents, you'll never run into speed-related constraints. You might start running into semantic constraints about limitations of the models' ability to differentiate different topics, but that's unlikely with domain-specific applications like you have.

Steadily making progress editing my first novel, and writing my second and third. Been averaging 6k words a week written for 3 months now. Remarkable how fast 25k words a month turns into books. Excited to get to the point of publishing. My first book is going through the royal road->kindle unlimited pipeline, and the initial reviews have been pretty damn good. I'm quite proud of it. I just need a title, been calling it 'Between Beast and Buddha: Book 1' in my head for ages now.

What sort of genre?

Xianxia, chinese cultivator fantasy. The second is book 2 of that same series, about an alcoholic monkey daoist loosely inspired by Sun Wukong, and the third is a weird sort of deconstruction of romantic tragedy set in fantasy post revolution not-france.

I won't say you're unoriginal.

TRON bike lighting update

Stuff finally came in from Ali and I got to work doing more proof-of-concept.

Firstly I found the batch of clear heatshrink tubing from Ali was too big. It never shrank down enough to provide the weather protection I needed for the WS2815 solder points, so I gave up and ordered something slightly smaller off of Amazon and then a bigger batch of it off of Ali to arrive at some future date. The smaller clear shrinkwrap tubing arrived 2 days later and worked well.

Of course I had to buy a new heat gun too since the one I was using died. And by that I mean melted itself. You have one job, Harbor Freight heat gun: don't melt yourself.

Then I got around to trying to control COB LED strips. These are much brighter and meant for visibility during the daytime rather than "art". I ended up with way too high gauge but-can't-hurt-right MOSFETs that, when I finally sat down and soldered them and put together like a half dozen different wires between the MOSFET, power source, breadboard, ESP32-C3 and COB LED actually worked fine the first time. I was able to turn them on and off entirely via software and then adjust to dimming them via software. Sweet.

I'm using white COB LEDs but I note they make RGB COB LEDs which seems like so, so much more wiring but the effect might be worth it. They're so bright! We'll see how the WS2815s handle the heavy lifting for now. I guess.

I think nothing really stands in the way of going into full production at this point. Production here meaning put WS2815 and COB LEDs on helmets and the cargo bike and seeing if I can coordinate them via a smartphone app via BLE. Then I guess branching out to vests and other accessories.

I am trying to decide between putting 18650 batteries on each helmet or just kinda having a bigger battery on the cargo bike that each helmet's logic board plugs into via barrel jack adapter. As long as they're not screw-tightened together it shouldn't be a strangle/whiplash risk.

Also I just randomly found these RGB LED backlights on Adafruit and I'm looking for an excuse to incorporate them.

Also, I'm really enjoying how tiny the ESP32-C3 chips are and something about them being RISC-V makes me feel like I'm finally getting the future I was promised.

Also! Were you aware you could just hide a whole accessible via WiFi system-on-chip in the head of a USB-C or Lightning cord? https://shop.hak5.org/products/omg-cable

Homies: Ride or Die update

Spent awhile trying to rotate the wheels on the car model while it was in motion.

The first problem was figuring out how to isolate the wheels from the model mesh. I was about to rig some scheme where I would take shots in the dark to come up with relative points in the centers of each wheel and then try to eyeball a definition of a cylinder that would fit them, then write some selector that would pick all of the submeshes contained inside of the cylinder and then I could apply transformation matrices just to those that will match the spin of the wheels at any given time. But then I realized the model file actually already identifies the wheels and I could just process the "group" directive. With the hard part done, it was a trivial but surprisingly annoying amount of transform math to translate (ha!) into code.

So, the wheels spin. Now I'm trying to get them to also turn with the steering. Should be able to use most of the same code, just rotating the wheels around a relative Y axis instead of a relative X axis, but of course I'm missing some detail and they barely move even if I have the car turning 30 degrees at a time.