Lead AI Engineer at Raumdeuter · based in Leipzig 🇩🇪 · previously at Mercedes‑Benz.

I build GDPR‑first, multi‑LLM production systems for German‑language media clients. Sentiment analysis, contextual question generation, learning loops that feed back into prompt design, self‑hosted GPU planning for when Cloud credits run out.

• Multi‑LLM adjudicator on Vertex AI — every /api/sentiment request fans out to two Gemini 3.x workers in parallel (Frankfurt‑pinned + EU multi‑region), reconciled by a Gemini judge. Cross‑cloud safety‑net fallback to Fireworks. Zero 2.5‑family models on the request path.
• Fragengenerator KI — admin tool that turns a Kicker‑style article into three fan‑facing prompts (text · ranking · poll) using an engineered German prompt, with a reveal‑edit UX, per‑card regenerate, and a mandatory "warum?" reason on every touch. The reasons + a supervisor Gemini's per‑generation critique feed the learning loop that improves the next prompt version.
• Sentiment analysis at scale — 5‑class buckets, 7‑class intents, 18‑topic taxonomy, register‑anchored German prompt v30 with disagreement guardrails. Live on Chemie Leipzig and the demo showcase.

Python / FastAPI · gRPC · PostgreSQL + MySQL · Docker Compose · LangChain / LangGraph · Vertex AI Gemini · Cloudflare R2 · Phoenix (OTel tracing) · Sentry · Hetzner · GCP

Not just Gemini — Fireworks-hosted GPT‑OSS as the cross‑cloud fallback, and a self‑hosted Llama 3.1 8B (llama.cpp on the EU box) for internal SLM work.

• Auditable data residency. Every LLM call in production runs in an EU region — Frankfurt or the eu multi‑region — with a paper trail. No global endpoints on the sentiment path. GDPR compliance is a design constraint, not a checklist.
• Learning from every request. Editing an LLM output should be a signal that goes somewhere useful. In the Fragengenerator flow, every admin edit + the reason for it lands in a structured table alongside a supervisor Gemini's grade of the original output — the corpus we mine for the next prompt.
• Runway‑aware infrastructure. Google Cloud credits are great until they aren't. Prototyping the Hetzner‑hosted SLM cutover now so the migration is a decision, not an emergency.

• Company: Raumdeuter GmbH
• Location: Leipzig, Germany
• Reach me through Raumdeuter or via GitHub.

_{Profile last refreshed 2026‑09‑21.}