Projects

Models and inference 5

  1. PYstylus-0.9B

    A model that draws like a hand

    0.93B parameters, trained on 1 rented GPU. It looks at a sketch and redraws it stroke by stroke, with human order, timing and handwritten labels. Paper included, weaknesses too.

  2. SHdsv4-flash-b300

    DeepSeek-V4-Flash on one Blackwell GPU

    An abliterated build packed into 1 RunPod image. Rent the GPU, start the image, and it answers on OpenAI- and Anthropic-style APIs.

  3. PYgptoss-spark

    gpt-oss-120b, 31% faster on the same box

    52.6 → 68.9 tok/s single-stream on one DGX Spark, from 6 patches and kernels built for its chip. The approaches that failed are written down too.

  4. s2pro-native

    Text to speech in C and CUDA, faster than realtime

    I threw away the Python stack around Fish Audio S2-Pro and rewrote the model in C. Wall RTF 0.51, first audio in 0.14 s, 33 voices, one binary.

  5. RSqwen3-tts-native

    Qwen3-TTS, talking in 94 ms

    The same rewrite in Rust and CUDA: 94 ms to first audio where stock SGLang needs 2.69 s, in 5.68 GB of GPU memory instead of 108.90 GB.

Reverse engineering 2

  1. Inkside

    I taught AI to write in Goodnotes

    Goodnotes keeps its ink in an undocumented file format, so I took it apart. Inkside writes study sheets as real, editable pen strokes: headings, maths, tables and diagrams.

  2. TScpe

    My 5G router's encrypted API, taken apart

    The router's web UI hides 348 endpoints behind RSA, PBKDF2, AES-256-CBC and HMAC. I read the crypto out of its own JavaScript and rebuilt it as 1 binary.

On the side 4

  1. TSweb-glass

    Liquid glass for the web, without melting the GPU

    34–40% GPU while glass moves, down from 72–90%, with 0 dropped frames and no WebGPU needed.

  2. downbeat

    One song on every phone in the room, at the same moment

    Phones join with a QR code and stay within ~0.5 ms of each other, with 0 late packets over a 4-minute song.

  3. RSbrowser-rs

    Browser automation for agents, at a third of the memory

    535 MB where Playwright MCP needs 1,496 MB. Pages come back as Markdown, steps can be batched, and a human can take over the live page.

  4. LGKA+

    The app my school checks before getting out of bed

    Native SwiftUI on iOS and Kotlin on Android, fed by a Cloudflare Worker that parses the school's PDFs once for everyone. 500+ of ~700 students open it every morning.

Proudest of 4

  1. PHPPholio

    A folder of Markdown in, a docs site out

    Static documentation with search, in plain PHP with 0 dependencies. Dairo's docs run on it.

  2. RSatlas

    My home server, end to end

    Wake-on-LAN from the Mac, my own photo library with local AI search, a SwiftUI app and a terminal in my pocket.

  3. school-ui

    The school's AI, inside the school

    A 120B model on one DGX Spark in the building, built by two of us: me and the school's head of IT. Nothing anyone types leaves the premises.

  4. Orin

    Your morning, read to you by two AI hosts

    Two hosts talk you through your inbox, calendar and news. It moved from Google Cloud to Rust on Cloudflare Workers; Orin Radio, a DJ for your Spotify, is an experiment on top.

Startup 1

  1. Kiste

    Linux machines that remember where you left off

    Firecracker microVMs on servers in Germany. Park one and it costs nothing; start it a day later and it is back in under 1 second, shells and servers still running. SSH, a desktop in the browser, forks, and a CLI that agents can drive.