AI Engineer | Data Analyst | RAG, LLM Agents, Analytics
3+ years across AI engineering at iiterate Technologies and fraud and renewal analytics at KPMG. Python, SQL, LangGraph, FastAPI, Power BI.
I'm an AI Engineer with an analyst's background. At iiterate Technologies I led a RAG chatbot for the municipality of Bad Breisig that residents use to find local government information. Before that I spent almost three years as a Business Analyst at KPMG, building fraud and renewal models on insurance data and the Power BI dashboards the team used to review claims. I care about the unglamorous parts: evaluation, data quality, and knowing when a model is wrong. I'm based in Mannheim and open to roles across Germany.
Residents of Bad Breisig needed an easy way to find local government information without digging through municipal websites and documents.
Led development and deployment of a RAG-based chatbot for the municipality, managing two interns from requirements to production rollout and iterating on real user query patterns.
In production for real residents.
Demand planners often have to accept forecasts from opaque models with no way to add their own judgment.
Built a multi-agent demand planning system at the Paretos Decision Hack in Heidelberg, where planners can override AI forecasts with reasoned justifications instead of accepting a black box.
3rd place. Improved forecast accuracy by 21.8% across 68,000+ item-months of data.
Spotting risky claims and likely renewals in insurance data.
Random Forest fraud model, Gradient Boosting renewal model, and Power BI dashboards for the claims team.
~20% better fraud model performance, ~30% less manual review effort.
Client work, no code or client data shown
People want quick, trustworthy answers to health questions without a model making things up.
An on-device assistant that answers only with extractive text pulled from source documents, no generated text, using encoder-based extractive QA and a cross-encoder reranker (ms-marco-MiniLM-L-6-v2). Designed for EU MDR Class I positioning.
Evaluated on a 50-question set.
Understanding and anticipating Deutsche Bahn train delays.
Built a Deutsche Bahn train delay intelligence platform, demoed live at an Accenture-hosted Claude Code community meetup in Mannheim and Heidelberg.
Teams need a RAG backend they can trust, not just one that sounds confident.
A RAG backend with a built-in evaluation framework to catch wrong answers before they reach users.
Response accuracy +25%, incorrect answers reaching users -30%.
Understanding how citizens make decisions during a crisis like a flood, to help improve civil protection coordination.
An agent-based evacuation simulation on a real street network (OSMnx) with LangGraph orchestrating each agent's decisions, visualized in Streamlit.
Generating clinically meaningful chest X-ray impressions automatically, without the model quietly ignoring the image.
A two-view ViT + BioBART architecture generating impressions from dual-view chest X-rays on the MIMIC-CXR dataset.
Clinical F1 about 0.51. Key finding: models can score well on standard metrics while ignoring the actual image (shortcut learning).
Manually sorting transactions into budget categories is tedious and easy to fall behind on.
A full-stack app with AI auto-categorization of transactions.
Cut manual sorting time by an estimated 70%.
Manually typing structured data out of scanned documents and images is slow and error-prone.
An OCR pipeline extracting and structuring text from images and PDFs, with image preprocessing to improve accuracy.
Accuracy +20% through image preprocessing, manual data entry time -35%.
Visually impaired users need real-time awareness of people and obstacles around them.
Built and deployed an object detection system for smart glasses with real-time person and obstacle recognition.
In production for a client.
Client work, code private
ML services in production need monitoring, or problems go unnoticed until users complain.
An ML service with monitoring and alerting on Kubernetes.
Incident detection time -40%, uptime +15%.
During a crisis, deciding who a limited number of responders reach first, and by which route.
A planner using A* for each request's route to safety and Dijkstra for a single responder's priority-ordered, multi-stop dispatch tour, on a real street network.
E-commerce brands need many on-brand product visuals, and reshoots or manual edits for every campaign variant are slow and expensive.
A modular ComfyUI workflow using ControlNet for pose consistency and an IPAdapter node for brand style transfer from a single source photo.
Cut visual variant production time by an estimated 70%, generating 10+ usable variations per image.
Generating new, usable fashion images from learned style patterns instead of sourcing new photography for every idea.
A GAN that generates new fashion images from learned style patterns.
About 80% of outputs were accepted as usable in an informal visual review (estimate).
Answering medical questions with grounded, explainable responses instead of an ungrounded chatbot.
A domain-aware medical RAG chatbot with vector retrieval over the MedQuAD dataset.
Making country-wise and athlete-level Olympic trends easy to explore.
An interactive Power BI dashboard for Olympic data storytelling.
Searching and answering questions from internal documents without manually digging through them.
A RAG application using LangChain and OpenAI embeddings with a Pinecone vector store.
Answering ad-hoc business requests quickly with SQL instead of waiting on a full reporting cycle.
SQL-based ad-hoc analysis answering 10+ business questions on sales trends, top channels and product performance.