AI Master's Student @ FAU · IT Specialist @ TUM
AI Platform & HPC Infrastructure Engineer
M.Sc. thesis at NHR@FAU on eBPF-based NFS rate limiting for HPC clusters
Core: Python · Linux · eBPF Platform: Kubernetes · Terraform Serving: FastAPI · Docker · Docker Compose Orchestration: Apache Airflow CI/CD: GitHub Actions · GHCR Monitoring: Prometheus · Grafana ML: PyTorch · XGBoost · scikit-learn · MLflow · Evidently LLM / RAG: LangChain · ChromaDB · Anthropic Claude
| Project | What it does | Stack |
|---|---|---|
| ai-platform-lab | End-to-end AI platform: Terraform-provisioned Kubernetes with remote state, monitoring and a CI/CD pipeline shipping containers to the cluster | Terraform, Kubernetes, GitHub Actions, Docker |
| gpu-scheduler | Fair-share GPU cluster scheduler with EASY backfill and starvation-free job aging | Python, FastAPI, Prometheus, Grafana, Docker |
| llm-inference-server | Continuous-batching inference server with iteration-level scheduling and TTFT/throughput metrics | PyTorch, Transformers, FastAPI, Prometheus, Docker |
| Project | What it does | Stack |
|---|---|---|
| pepper-games | Voice games, tablet kiosk and web admin panel for a SoftBank Pepper robot, answered by a self-hosted LLM | Python, naoqi SDK, Whisper, gpt-oss-120b |
One shared stack across the series, where each project adds a piece the previous one did not have.
| Project | What it does | Stack |
|---|---|---|
| churn-prediction | The baseline of the series: a model registry and drift detection wired into a serving API | XGBoost, MLflow, Evidently, Prometheus, Docker, K8s |
| price-prediction | Same pipeline pointed at a regression target, so the metrics are errors rather than class scores | XGBoost, MLflow, Evidently, FastAPI, Docker |
| anomaly-detection | Drops the labels entirely and scores sensor readings with an unsupervised model | IsolationForest, MLflow, Evidently, FastAPI, Docker |
| airflow-ml-pipeline | Takes training off my machine into a scheduled DAG that retrains when drift crosses a threshold | Airflow, MLflow, Docker Compose |
| image-classifier | First GPU-trained deep learning model here, a CIFAR-10 CNN behind the same API shape | PyTorch, MLflow, FastAPI, Docker |
| document-qa | Answers from retrieved documents instead of a trained model, and evaluates the retrieval itself | LangChain, ChromaDB, Claude, FastAPI, Docker |
| mlops-sentiment-pipeline | Splits the single container into a multi-service Compose stack with a transformer doing inference | DistilBERT, MLflow, Evidently, FastAPI, Docker Compose |
Every MLOps and Infrastructure project ships with tests, a Dockerfile, and a GitHub Actions pipeline.