ML platform / LLM engineer, software engineer and data scientist in Bielefeld, Germany. I build and run LLM platforms: model selection, GPU serving on Kubernetes, RAG, agentic workflows and the evaluation that tells you whether any of it works.
Currently at mittwald, AI Hosting. M.Sc. Data Science (Hochschule Bielefeld), B.Sc. Economics (Universität Bielefeld).
LinkedIn · Publications · mittwald AI Hosting docs
- LLM inference on 2x NVIDIA RTX PRO 6000 Blackwell with vLLM, Kubernetes, Helm, KEDA and the NVIDIA KAI scheduler
- Model portfolio and selection for mittwald AI Hosting, from Qwen3.5-0.8B for routing to 122B models, documented in the developer docs
- Evaluation, load testing and guardrails for production LLM APIs
- Agents with tool calling (TypeScript, Mastra), RAG building blocks (embeddings, reranker, OCR) and cascade routing from the smallest to the largest model
- Coding agents (Claude Code, Codex) against self-hosted models, every change reviewed
- Fischer, D. et al. (2024). Demystifying Reinforcement Learning in Production Scheduling via Explainable AI. arXiv:2408.09841
- Schwede, C., Fischer, D. (2024). Learning Simulation-Based Digital Twins for Discrete Material Flow Systems: A Review. Winter Simulation Conference 2024. doi:10.1109/WSC63780.2024.10838729
| Project | What it is |
|---|---|
| awesome-vllm-plugins | Curated list of open vLLM plugins (vLLM plugin system) |
| simpleaudiospectral | Self-hosted signal analysis web app; releases, CI, coverage, SBOM and SLSA provenance, multi-arch images |
| rustypastechat | Android chat app (Kotlin, Material 3) with an assistant on any OpenAI-compatible endpoint; tested, released |
| resnet-bilstm-attention-dt | Master's thesis code: verification and validation of simulation-based digital twins |
| papers | PDFs of my publications |
Two Linux servers running about 130 containerised services with Docker Compose, kept as infrastructure as code: VictoriaMetrics, Grafana and Alertmanager for monitoring, SOPS-encrypted secrets, access through Tailscale, encrypted off-site backups, around 800 unit tests and pre-commit hooks. Document OCR with self-hosted models, an AI agent with long-term memory, and a Home Assistant smart home the agent can control through MCP. The repository is private; happy to walk through it.