AI engineer in Ilmenau, Germany. I build production systems for search, question answering, speech and document understanding. I care about turning messy files like invoices and PDFs into clean data and about shipping services that are fast, stable and easy to observe.
Email • LinkedIn • Location: Ilmenau, DE
- LLMs and retrieval with grounded answers, citations and simple chunking that fits the data
- Speech and document AI with Whisper, Wav2Vec2, OCR and layout aware models for invoices and tables
- Production work with FastAPI, Docker, CI and vLLM along with quantization, batching and streaming
- Safety and evaluation with PII scrubbing, refusal rules, golden sets, EM and F1, RAGAS, A B tests and tracing
- Clinical assistant pilots cut lookup time from minutes to seconds and reduced post visit documentation by about 25 to 35 percent
- Medical speech models improved word error rate by 6 to 8 points and kept real time behavior stable
- Invoice and document extraction reached over 90 percent F1 on vendor, VAT, IBAN and totals and lowered manual review
- Serving became about 30 percent faster at the median by using lower precision models and token streaming
- Shift AI - LLM powered shift scheduling with constraint aware planning and a FastAPI backend
- AI Agents Automation - Multi agent system that files and routes work, posts updates to chat and looks up fixes from a knowledge base
- Early Dyslexia Detection - Multimodal screening that combines handwriting analysis and speech features
- Invoice DocAI OCR and layout models with an LLM fallback for hard cases plus batching and caching
- Clinical Consultation Assistant grounded answers with citations on Vertex AI and Cloud Run with PHI scrubbing
Python, C++ • FastAPI • LangChain, LangGraph, LlamaIndex • vLLM • FAISS, pgvector •
Whisper, Wav2Vec2 • Tesseract, DocTR, LayoutLMv3, Donut • GCP, Vertex AI • PostgreSQL • Docker, CI • Git, Jira
I value readable code, clear APIs and measurable impact. I ship in small steps, watch the graphs and iterate. Open to collaboration and feedback. Issues and PRs are welcome.