Skip to content
View invaliddayta's full-sized avatar
2 followers · 3 following

Block or report invaliddayta

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
invaliddayta/README.md

hi, i'm paul

embedded linux and edge ai engineer in vienna, austria. secure, reproducible, fast, and boring in production

  • cto at STØK (company in formation): real-time stocksport scoring from cameras on a jetson running nixos. that code is closed source
  • before that at agilox: moved a fleet of 3,000+ autonomous robots to measured boot, full disk encryption and secure boot, plus the tpm provisioning and ota updates behind it
  • msc in ai at jku linz

what's here

  • llama.cpp-opt: llama.cpp fork with my own cuda kernels. a 27b model on one rtx 3090 verifies 8 draft tokens in 33 ms instead of 54, output bit-identical
  • llm-opt: the benchmarks and notes behind it
  • nixos-orangepi-zero2w and nixos-rockpi-s: nixos on tiny arm boards. one nix build, bootable sd card
  • agent-perch: a small browser terminal for coding agents, on top of tmux
  • dsh-tui-nix: slim nix package for the deepseek harness tui

work with me

open to freelance projects in embedded linux, device security and edge ai. on-site in vienna or linz, or remote. linkedin or paul@protom.eu

Pinned Loading

  1. llama.cpp-opt llama.cpp-opt Public

    llama.cpp fork tuned for speculative decoding on sm_86: int8 small-batch GEMM, fused q4 KV attention, GPU tool grammar, DFlash2

    C++

  2. llm-opt llm-opt Public

    Making a 27B model decode fast on one RTX 3090: kernels, benchmarks and measurements for llama.cpp speculative decoding

    Python

  3. nixos-orangepi-zero2w nixos-orangepi-zero2w Public

    NixOS board support for the Orange Pi Zero 2W

    Nix 1

  4. nixos-rockpi-s nixos-rockpi-s Public

    A fully automated, plug-and-play NixOS flake for the Radxa Rock Pi S (RK3308). Builds a bootable SD image with U-Boot, RAMdisk, and serial console fixes out of the box.

    Nix 3 1

  5. agent-perch agent-perch Public

    A small, hackable browser terminal for your coding agents. OpenCode, Codex, DeepSeek Harness. Your machine. Your tmux sessions.

    TypeScript 1 1

  6. s90c-tools s90c-tools Public

    Root commands, screenshots and app cleanup for Samsung S90C. Tested on one Tizen 7 build. Linux/Nix.

    Python 5 1