aspiring cyborg, trying to fix the limited bandwidth of humans
cto, linux and edge ai engineer. i like building robots
- STØK: stocksport is still scored with a tape measure. in 2026. we do it with four cameras and a jetson running nixos instead. that code is closed, sorry
- before that at agilox i moved 3,000+ autonomous robots onto measured boot, full disk encryption and secure boot, plus the tpm provisioning and ota updates behind it
- bsc in ai at jku linz, now doing the msc there
- llama.cpp-opt: not everyone has a datacenter. i have one rtx 3090 hanging off a usb4 cable. my cuda kernels take a 27b model's verify step from 54 ms to 33. if the output hashes don't match, it doesn't ship
- llm-opt: the receipts. benchmarks, profiles and everything that didn't work
- samsung-tv-tools: my samsung tv kept nagging me with update popups and apps i never asked for. so i rooted it
- nixos-orangepi-zero2w and nixos-rockpi-s: i keep putting nixos on tiny arm boards and nobody can stop me
- agent-perch: a small browser terminal for my coding agents, on top of tmux. so i can check on them from my phone
- llama.cpp: a shared-memory race in the cuda flash attention kernel. 2.2 million racecheck hazards down to zero (#30573)
- omarchy-mlx: 4-bit group 32 and 8-bit models were stuck on a slow fallback kernel. on my macbook air running linux a 7b model now reads prompts ~8x faster (#179)
freelance embedded linux, device security and edge ai. vienna, linz or remote. linkedin or [email protected]