About
I am an IT professional with a background in computer engineering, artificial intelligence, research computing platforms, and secure network infrastructure. I currently manage IT infrastructure and high-performance computing environments across several departments at the Technical University of Munich. My goal is to combine IT infrastructure, software systems, and AI to develop reliable, scalable platforms that address real-world challenges.
Alongside this I am completing an M.Sc. in Artificial Intelligence at FAU Erlangen-Nuremberg.
What I work on
I run production GPU and AI infrastructure at TUM: job queueing and fair-use allocation on a shared GPU cluster, model serving, a self-hosted large-language-model deployment, and a LiteLLM gateway in front of the GPT API. Around that sits the monitoring, the server and network architecture, and the automation that keeps it usable for the researchers on it.
Before this I worked as a computer engineer on network and infrastructure projects, and spent a year between IT and product at a startup in Berlin.
Projects
End-to-end internal AI platform: an LLM gateway with observability, provisioned and deployed entirely through infrastructure as code.
- LiteLLM gateway in front of a locally hosted model, with master-key auth from the environment and per-deployment rate limits.
- Terraform for the AWS side: an S3 bucket with encryption and versioning, and remote state with native locking.
- Flask service containerized non-root and run on Kubernetes with probes and resource limits, Prometheus and Grafana scraping both.
- CI checks the Terraform, Docker, Kubernetes and Python layers on every push.
Continuous-batching inference server built from scratch over Hugging Face Transformers, with iteration-level scheduling and TTFT/throughput metrics.
Job scheduler for a shared GPU cluster: fair-share priority, EASY backfill and starvation-free job aging, with Prometheus metrics.
Voice-driven games, a tablet kiosk and a web admin panel for a SoftBank Pepper robot, with answers generated by a self-hosted LLM.
Containerized REST inference service serving a DistilBERT sentiment model, deployed live on Hugging Face Spaces.
XGBoost churn service with a model registry, Evidently drift detection, Prometheus/Grafana monitoring and Kubernetes manifests.
Retrieval-augmented Q&A over uploaded documents with source citations and a recall@k retrieval evaluation gate in CI.
Scheduled end-to-end training pipeline with a quality gate and a drift monitor (KS test) that triggers automatic retraining.
Sentiment pipeline with experiment tracking, drift detection and a two-service Docker Compose stack.
California house-price regression service with experiment tracking and drift reporting.
Unsupervised anomaly detection over sensor telemetry, served as an API with tracking and drift reporting.
CIFAR-10 convolutional network served over an inference API, with training tracked in MLflow.
Every project ships with tests, a GitHub Actions pipeline and a container it runs in. More on GitHub.
Skills
- AI / ML
- LLM deployment and self-hosting, OpenAI-compatible LLM gateways (LiteLLM), OpenAI / GPT API integration, Hugging Face Transformers, PyTorch, model serving and inference APIs (FastAPI), speech recognition and speaker diarization, classical ML
- Infrastructure & MLOps
- GPU cluster operations, Kubernetes, Docker, Terraform (remote state on an S3 backend), AWS (S3), GitHub Actions and CI / CD, Prometheus and Grafana monitoring, VMware, Hyper-V, Linux / Windows / macOS administration, network architecture, storage configuration, deployment / DevOps workflows
- Programming
- Python, Java, C / C++, Bash; backend services (Flask, FastAPI)
- Other
- Secure IT architecture and network hardening, technical team leadership