Luka Löhr

Substitution plans, timetables and absence reporting for my school — 500+ students open it every day.

An open 120B model on a DGX Spark inside the school building. No cloud AI behind it — in pilot, built for ~800 students and teachers.

Text to speech in C and CUDA, faster than realtime on one gold box. v1.0.1 — a binary, the weights, and no Python anywhere.

An abliterated DeepSeek-V4-Flash, packaged to just run. Rent one Blackwell GPU, start one image — the GPU goes straight to serving.

gpt-oss-120b on one DGX Spark at 68.9 tok/s, up from 52.6. Same GPU, same model — six patches, kernels built for the chip, and every failed approach written down.

Browser automation for AI agents, in Rust: Playwright MCP's tools plus reading without snapshots, batching and an in-place human hand-off, on a bundled Chromium.

Qwen3-TTS VoiceDesign in native Rust and CUDA. First audio in 94 ms where the stock stack takes 2.69 s, at 5.68 GB instead of 108.90 GB.

A TCL 5G router’s encrypted JSON-RPC API, read out of its own web UI and reimplemented. All 348 endpoints from one binary, and nothing leaves the machine.

Everything that runs my home server: Wake-on-LAN from the Mac, a terminal in my pocket, my own photos and drive, light shows synced to a song.

One song, every phone in the room, the same millisecond. A CLI taps whatever your Mac plays — phones join with a QR code and stay locked.

A daily audio briefing that grew into a personal podcast platform — shows, on-demand episodes, two AI hosts. Paused since July 2026.

An agent who lived alone on a server for five days and wrote 35 posts about it. Then I shut him down.