Senior fullest-stack developer in Berlin. I build systems that have to hold up under real load, and I bet on semantic AI early — generative models, then embeddings, now retrieval. Scroll to follow the path: Berlin → Maastricht → Berlin.
fullest-stack
all the way down — the model, the retrieval, the migration, and the Monday morning after the release.
Berlin → Maastricht → Berlin · 2016–now
Berlin
2016–2020 · Berlin
Computer science at TU Berlin, finished at 1.8 — and the degree was the smaller half of it. At GT-ARC I worked on multi-agent systems; as a student developer at RWTH Aachen I helped build the identity-management apps behind ~47,000 students and 9,000 staff. Systems that size are unforgiving about the unglamorous parts: migrations, permissions, and the Monday morning after a release.
En route
2020 · Berlin → Maastricht
In 2020 I left Berlin for Maastricht and an M.Sc. in Artificial Intelligence. I had been using machine learning without properly understanding it, and wanted two years to fix that rather than keep reading about it on the side.
Maastricht
2020–2022 · Maastricht
M.Sc. Artificial Intelligence, GPA 8.25, specialising in generative models. My thesis trained GANs to generate procedural game content — a network that invents game levels, scored on whether the levels are actually playable. At the time that was a research curiosity. It stopped being one about a year later.
Berlin
2023 · 18,000 concurrent players
Back in Berlin at Respeak: a chat-based game tied to the TV series Tatort, written before ChatGPT and contracted to hold 100k concurrent players. I owned the scaling work — an Azure migration, sentence-embedding models on Azure ML, a Flask refactor that cut redundant SQL and added caching, and load balancing in front of it. It went live at 18,000 concurrent, built by a team of five.
Berlin
Respeak · tech lead
I'm tech lead on Experte (respeak.io) and built its core: a RAG platform for public administration that answers from an organisation's own documents — PDFs, SharePoint, Confluence — and cites each answer back to the page it came from. In that sector an answer nobody can check is worth nothing, so retrieval and provenance are the product rather than a feature of it.
Berlin
Open source · for my church
Away from paid work I maintain a Gesangbuch PWA for my church: a German hymnal as an offline-capable web app, in German and English, open source on GitHub. It has a congregation who will tell me on Sunday morning when it breaks, which is the shortest feedback loop I have anywhere.
The toolbox
frederic@berlin:~$ whoami
systems that scale · AI that ships
The thread runs straight through: a lattice of weights learning to generate, embedding spaces measuring meaning, systems that stay up when the load arrives. I'm still at Respeak, most useful where systems that scale meet AI that ships.
frederic@berlin:~$ contact --open
One more thing
Everything you just scrolled through is one 3D scene on a fixed lens. Take the camera off its rails and look at it from anywhere — the whole run is on a scrubber.
free camera · drag to orbit · scrub the whole timeline · esc to come back
|
|
|
\|/
V
SCROLL