bavaria, germany — applying to study computer science
I’m Robert — I build AI‑assisted software end to end, keep a few small machines running at home, and ship things I actually use. Abitur 2026, grade 1.7.
And yes, I actually do like em dashes. ;)
01 — built with models
I develop with AI coding agents every day and I’m open about it: I scope the tasks and own the results. This is my own local tracking across my machines — not an estimate.
tokens · august 24 – october 4, 2026 · ≈ $5,121 api‑equivalent
codex · 2.72B · 28% · mostly gpt-6-astra claude code · 6.92B · 72% · opus-5-5
updated october 5, 2026
02 — the work
My projects — the list updates as I build. Most are public, some stay private. I make things that actually benefit me, plus hobby projects I just think are cool — often deliberately oversized, to push AI‑assisted development to its limits.
A Bun monorepo — SvelteKit 2 / Svelte 5 in front, a Rust (Axum) workspace staged behind — where every page is prerendered from Markdown and YAML: no database, no API on the content path. Strict hash‑mode CSP, a hand‑rolled bilingual layer, and an Ask assistant streaming over server‑sent events; chat history stays in your browser.
Remote control for the machines on my desk, from the Mac menu bar or a phone. It shows which system a dual‑boot machine is running, wakes and sleeps it, reboots it into its other OS, and updates or restarts the services on it — every action one ssh round trip, no daemon, no open port. Built for the Legion 7 (CachyOS / Windows 11) and T3 Code, now config‑driven for any machine.
A self‑driving agent for Trackmania Nations Forever that sees only pixels — no game‑engine internals. Reinforcement learning over a pool of about 10,000 tracks, so the policy learns driving in general rather than one map; the benchmark is a gold medal within 50 attempts on tracks it has never seen. Trains on the Legion 7.
A playful, magnetic model‑and‑reasoning picker for ChatGPT. It opens from the native composer selector and delegates every choice to ChatGPT’s own controls: a draggable puck that snaps between exact stops, compact and full layouts, synthesized clicks, and a fallback to the native picker whenever fewer choices are available.
An instrument stage and music studio for macOS: eleven fully synthesized instruments, a track editor with lanes, a drum step grid, velocity, loops and undo, playlists, and export to WAV, AIFF, ALAC or AAC at up to 96 kHz. The stage renders the music as light while you play.
also public · music fetch · saf · everything else on github
03 — the story
I’m from a small town in Bavaria; school was a boarding school on lake Ammersee. My first projects, around year 9, were small and scrappy — a Python face tracker built on an AI vision model, game prototypes in Unreal Engine 5, Unity, and Godot. The projects grew up with me.
The most reliable way I learn systems is running them myself. It started with a Windows tower and a hardware phase — overclocking, file systems, PC building — until the interest tipped into software: Arch Linux on an old Lenovo IdeaPad as the very first distro, no training wheels, then a rabbit hole of distributions. Today it’s a small network: a tower PC on the desk as the daily driver at home, the MacBook Pro for university and everything mobile, a Raspberry Pi that runs nothing itself but orchestrates the rest around the clock, and the Legion 7 as an always‑on remote machine — Moonlight, small training runs, switched between CachyOS and Windows from a menu bar. The phone fills the gaps in between.
For four years I ran sound, lighting, and stage tech for my school’s events — the weekly assembly plus everything with an audience. The rest of the record: game master of a pen‑and‑paper group, student council, Model United Nations in Copenhagen, and two history awards I didn’t see coming.
04 — writing
Nothing here yet — honestly. The blog ships with the site and the pipeline is tested; the words are taking a little longer than the code. When the first note lands, rss.xml will have it first.
05 · the fastest way is email
Everything else — the CV in English and German, GitHub, RSS — lives in one place: all links →
Ask about the work, the tokens, the story. Answers come from the site’s own content.