Machine Learning Software Engineer
I build production-oriented AI systems across LLMs, retrieval, embeddings, GraphQL agents, real-time inference, and multimodal applications — from fine-tuned models on Hugging Face to agent tooling in Rust.
Open-source contributor to llama.cpp and the GraphQL AI Working Group.
Models, agent tooling and research I've shipped recently.
Fine-tuned Qwen3-Embedding-0.6B to map natural-language questions onto GraphQL Type.field coordinates, improving owner-type disambiguation across large schemas. Published on Hugging Face with SentenceTransformer weights and GGUF builds for local inference.
In the GraphQL AI Working Group benchmark: 63.6% field recall@50, ahead of OpenAI text-embedding-3-large (59.8%).
A Text2SQL text-diffusion LLM, including training code and a playground for step-by-step inference.
GraphQL MCP server enabling dynamic, elastic type discovery with type dependency trees and optimized discovery pipelines, designed for advanced agent-driven workflows.
Real-time AI data agent: upload a CSV and talk to your live analyst. Streams live graphs, voice results, and transcripts, and runs complex tasks like stance annotation while updating Google Sheets in real time.
1st place, Berlin AI Hackathon.
Explore cities with a real-time, AI-generated tour guide that tells stories, builds routes, and streams audio.
Autoencoder-based "neuron surgery" on LLMs: extract and amplify neuron features linked to specific topics (e.g. SQL generation), enabling targeted personality shifts by boosting activations at inference time.
Small, on-device CNN image classification in Flutter via ONNX.
Physical ray-tracing simulation for curved optical fibers in MATLAB.
Find similar recipes from a picture of a dish.
Play chess just by thinking about the moves with a consumer-grade EEG. Works by thinking of binary start and end positions of moves, translated to chess coordinates.
Trained a stance detection model and built a platform that analyzes top news headlines and Twitter posts in real time, clustering topics and sentiment.
Optical measurement of algae contamination and oil pollution, plus water flow rate, salt concentration, and temperature, transmitted via LoRaWAN from the North Sea.
3D molecule viewer and quizzes: play against friends over the internet, or via Bluetooth when no connection is available.
Contributions to projects and communities shaping AI tooling.
GraphQL Foundation
Explores how GraphQL can best support the growing needs of AI and machine learning. Working group participant since October 2025.
ggml-org
The widely used open-source engine for running LLMs locally in C/C++.
Talking about AI with fellow engineers.
Der AI Podcast für Software Engineers · with Thore & Leon
A podcast where two senior software engineers discuss what is happening in AI: new models, tools and companies, and what they mean for people who build software.
Listen on SpotifyPeer-reviewed publications.
Ranno, Manjunatha, Koritzius, et al. · Scientific Reports
A patient-specific computational model coupling hemodynamics, drug kinetics, and tissue growth to investigate in-stent restenosis after drug-eluting stent implantation.
Protze, Orland, Haldar, Koritzius, Terboven · Springer
Originating from my bachelor thesis, introduces a tool to accurately separate and quantify model factors in layered parallel programming paradigms for Hybrid MPI+OpenMP metric, providing deeper insights into sources of inefficiency—such as load imbalance and synchronization overhead—across applications with various concurrency patterns.