~/

Magic Traces

Tue, 05 May 2026 00:00:00 +0000

Generate coloring book style images that you can trace from the screen with pen and paper. You can generate images from a photo or voice description. You can use it from the browser or install it as a PWA. It even has local history which is synced across devices.

When my daughter, after visiting a museum with her kindergarten, told me that there was a display, and you can pick the image, place a sheet of paper above and trace the image. It was surprisingly fun for 6-year-olds. I thought, but what if she can pick any image she wants with a voice command? Or better transform a photo into a drawing? So we vibecoded the whole shit with my wife. Sure, it didn’t work properly, and I had to fix the backend manually, but it was a fun bonding experience. And I’m proud of it at this point, it’s a nice small feature complete product.

Smarter RSS Filter

Mon, 04 May 2026 00:00:00 +0000

Filter RSS feed with LLM. You can specify topics or content style to include or exclude. It is useful for being more focused on what you care about and avoiding noise. For me personally, it helps to avoid AI FOMO and clickbait.

I switched to RSS reader for my personal news feed a while ago. It’s pretty refreshing to use hronologial feed with only content that I care about. The problem is some of the sources are noisy but still useful. So I made a simple RSS proxy that filters topics that you specify. It’s LLM-based, so it’s a bit smarter than regexp and can detect fuzzy still like clickbait, rageabait. Or, for example, include only news about local LLMs and exclude OpenAI drama.

Training of useful small language models.

Mon, 04 May 2026 00:00:00 +0000

How full line complete works in PyCharm and other JetBrains IDEs

Paper explains how a tiny (100M parameters) code completion model was trained. I’m fascinated by SLMs (small language models), and it’s always a joy to find when they become more than just toys but useful in the products. You don’t need any math or deep learning knowledge to read this paper, it explains the setup and trainig data in plain english most of it will be accessible for an engineer.

Inroduction to cybernetics and Stafford Beer ideas.

Sun, 03 May 2026 00:00:00 +0000

The unaccountability machine is a great book as an intro to Stafford Beer ideas on cybernetics and how it could be applied to organizations. Together with Brain of the Firm and The Human Use of Human Beings episodes of GIU podcast it entertained me during the fall 2025.

Peak of Stafford’s ideas is project Cybersyn. Imagine the new Chilean government building a nationwide cybernetics network for distributed descision-making in the 70th before the Internet era. It’s pure sci-fi in real life. There is also a great The Santiago Boys podcast from Evgeny Morozov.

Recasing of the Volca Modular to banana synth format.

Fri, 01 May 2026 00:00:00 +0000

Volca Modular is a poor man’s Buchla Music Easel, a modular synthesizer released by Korg. It’s a fun little synth with fiddly controls and simple but powerful sequencer. At some point I was annoyed by controls fiddliness enough that I decided to build a new better interface for it.

Bigger knobs and patch cables were obvious, but the challenging part was to keep the keyboard fully functional to be able to use it and the sequencer. I ended up with the idea to build a laser-etched plywood case with a precise slot for the keyboard. To my surprise, it worked right from the first try. I also added attenuators to oscillator CV ins (I missed it so much) and dropped a few extras like clock dividers, euclidean sequencers, and R2R DAC.