AI Researcher & Engineer
I build measurable, reliable RAG systems and applied machine-learning tools for real-world use.
Master's student in Applied Research in Engineering Sciences at Ansbach University of Applied Sciences and working student in AI at Bechtle Systemhaus. My work spans RAG benchmarking, multilingual technical retrieval, and multimodal learning.
Research on evaluating complete RAG pipelines, multilingual retrieval, and production-ready AI systems.
NGEN-AI Conference · Trento, Italy · 2026
Presents a modular benchmarking framework that attaches to existing Retrieval-Augmented Generation (RAG) pipelines as a non-invasive evaluation layer. It sweeps the full configuration space — chunking, retrieval, prompting, and generation models — combining LLM-based answer-quality scores, retrieval correctness measures, and runtime indicators.
ARC — Ansbach Research Conference · Hochschule Ansbach · 2026
Poster presentation summarising why complete RAG configurations should be benchmarked rather than isolated components: prompt template, retrieval strategy, model choice, and metric choice each shift which configuration looks best.
Bachelor Thesis · Hochschule Ansbach · 2026
Develops a RAG system for multilingual information retrieval from ~1,500 technical product datasheets: PyPDF extraction, hybrid vector store (FAISS + Chroma), LangChain orchestration, and GPT-5 mini generation. Evaluated with retrieval metrics (Recall@k, MRR) and answer metrics (F1, accuracy), showing significantly better precision and more stable multilingual answers than pure prompting.
Research implementations first, followed by selected applied software projects.
Sports AI · Hackathon
Player pose classification from media-day photos in a team of three: CNNs and a YOLO-based object detection model combined with MediaPipe pose estimation to classify players by body posture, visible body parts, arm position, and scene objects (e.g., a football). Reached 98% validation and 97% test accuracy and was invited to the on-site final round at the Eintracht Tech Lab in Frankfurt.
PythonCNNYOLOMediaPipe
Clinical AI
Research codebase for multimodal physiological signals: respiratory-sound classification (ICBHI 2017) with CNN/GRU models and leave-one-subject-out evaluation, plus an ECG/wearable pipeline with signal-quality assessment — packaged with a CLI, test suite, and MLflow experiment tracking.
Pythonscikit-learnXGBoostMLflow
Clinical AI
Multimodal patient-risk prediction (24–48h deterioration) on synthetic ICU time-series. Compares MLP, LSTM, and Transformer architectures with proper clinical evaluation: AUROC/AUPRC, calibration, and threshold analysis, plus a CLI training loop and Streamlit demo.
PythonPyTorchOptunaStreamlit
RAG Systems
Production-grade RAG system for curating arXiv papers: Airflow ingestion pipelines, hybrid BM25 + dense retrieval on OpenSearch, FastAPI service, and Streamlit/Gradio frontends — all Dockerized.
PythonFastAPIOpenSearchAirflowDocker
Applied Software
Full-stack tool that calculates shipping and export costs from product dimensions, pallet sizes, and shipping-house price lists — accessible via a Telegram bot and a web dashboard.
FastAPINext.jsTypeScriptSQLAlchemy
Applied Software
Personal finance application in a Turborepo monorepo: transaction and category management, dashboard, and financial-health insights backed by Supabase, with an end-to-end test suite.
TypeScriptTurborepoSupabaseVitest
LLM Tooling
Platform for benchmarking LLMs and SLMs — especially local models on consumer hardware — across tasks and metrics (exact match, multiple choice). Evaluates via any OpenAI-compatible server (Ollama, vLLM, LM Studio) or in-process Hugging Face models, stores all runs in SQLite, and ranks models on a leaderboard with per-sample drill-down.
PythonFastAPIReactStreamlitSQLite
LLM Tooling
Interactive playground for seeing how sampling parameters (temperature, top-k, top-p, repetition penalty) shape autoregressive output. Compares open models from GPT-2 (2019) to LFM2.5 (2026), visualizes per-token probability charts, and includes a CLI for headless experiments and side-by-side model comparisons.
PythonStreamlitTransformersCLI
Applied AI · Bechtle
Web-scraping and analysis tool for AI, VR, and AR news from sources like arXiv, The Verge, and Scinexx. Automatically collects articles, generates summaries with OpenAI's GPT-3.5 Turbo, and provides structured data and visualizations via a frontend for research purposes.
PythonGPT-3.5 TurboStreamlitWeb Scraping
Computer Vision
ML algorithm for detecting failed prints in 3D printing. Compared three approaches: template matching (84% accuracy), a fine-tuned YOLOv8n model (100%), and a custom-built CNN (99%).
PythonYOLOv8CNNOpenCV
Computer Vision
Disease detection for potato and tomato plants from leaf images, built with two approaches: a hand-programmed CNN and a fine-tuned, pre-trained YOLOv8n model.
PythonCNNYOLOv8n
Robotics
Programmed an Alphabot2 robot to autonomously escape a labyrinth using its line-tracking functionality.
PythonRoboticsLine Tracking
Competitions, course certificates, and statements of accomplishment.
Bundesliga · DFL · Eintracht Tech · AWS · Slalom
Reached the on-site final round at the Eintracht Tech Lab in Frankfurt after excelling in the multi-week competition challenge set by EintrachtTech and DFL.
View certificateHugging Face Agents Course
Certificate of Achievement for successfully completing the Foundations of Agents unit of the Hugging Face Agents Course.
View certificateDataCamp · Statement of Accomplishment
Completed DataCamp's course on building multi-agent systems with LangGraph.
View certificateDataCamp · Statement of Accomplishment
Completed DataCamp's course on supervised machine learning with scikit-learn.
View certificateDataCamp · Statement of Accomplishment
Completed DataCamp's course on unsupervised learning - clustering, dimensionality reduction, and more - in Python.
View certificateApplied AI experience and the technical toolkit behind my research and engineering work.
Feb 2025 — Present
Bechtle Systemhaus Holding AG · Remote
AI research and development, including an automated market-intelligence tool for collecting, summarising, and presenting AI, VR, and AR news.
Sep 2024 — Jan 2025
ANregiomed gKU · Ansbach
Supported the implementation of a patient data management system, working across clinical software, user interfaces, and server infrastructure.
Oct 2022 — Mar 2025
ANregiomed gKU · Ansbach
A concise overview, with the complete document available as a PDF.
Happy to talk about research, collaborations, or interesting problems.