Hi, I'm Leon.

The research project I spent the most time on. The detector sits inside the autoregressive decoding loop and scores every candidate token against the retrieved context, so an unsupported claim is penalised before it is ever sampled instead of being caught after the fact. The number detector works well and cuts numeric hallucinations substantially at limited overhead. The factual detector is far more experimental and stands or falls with tokenizer alignment and the latency you are willing to spend. I find the direction promising, but its limitations are real and worth naming.

• Python
• PyTorch
• Transformers
• RAG
• LLM evaluation

Photo: © Claudia Poppe

This was one of the most meaningful projects I helped turn from a hackathon idea into a funded product. We digitalised how street newspapers were sold by enabling vendors to accept digital payments through QR codes, removing the need for a bank account or a mobile phone. I led the project as a project manager and built the backend architecture by implementing the REST API and the security layer for the payments. The project combined product thinking, systems design and practical delivery in a way that had a real public impact.

• Go
• REST APIs
• SQL
• Vue.js
• Payments

This was the interdisciplinary project of my Data Science master, where the point is to work with a faculty outside computer science. I chose the juridical faculty and built a retrieval system over the Viennese building regulations with a fellow student. The corpus was far too specific for any general purpose model (at the time). The answers had to be in German and everything had to run locally, without LLM providers or frameworks like LangChain. Therefore, we implemented chunking, retrieval and reranking ourselves around Elasticsearch and Ollama. What shocked me most wasn't the finding of the higher importance of the chunking strategy over the embedding model but that a simple keyword-based evaluation revealed a non-deterministic behaviour in our retrieval pipeline.

• Python
• FastAPI
• Elasticsearch
• Ollama
• Docker
• RAG

Together with a fellow student I deployed a demo shop split into eleven microservices on Google Kubernetes Engine and then pushed it until it broke. We automated the load generator with Terraform and Ansible, watched the cluster through Prometheus and Grafana and saw the frontend and the currency service give out at around a thousand concurrent users. Rebuilding the same setup with K3s afterwards showed how much of Kubernetes a managed service quietly hides from you.

• Kubernetes
• Terraform
• Ansible
• Prometheus
• Google Cloud

More than 500 signatures on a petition to close a short stretch of road to cars to create a vivid place for the neighborhood and for the elementary school next to it. It started as a simple idea by inviting neighbours to share dinner on the street in summer 2025. One year later it grew into a small civic movement once we realised how tangibly the experience strengthened community life. The outcome of the petition is still open but the initiative continues to bring people together and might even lead to a community yard at the very same place. This is one of my favourite personal projects.