Thore Koritzius
ML Founding EngineerMunich, Germany
I build ML systems around LLMs, retrieval and embedding models, from training and fine-tuning models to making them run fast and reliably in production. I studied computer science at RWTH Aachen.
On the side I contribute to llama.cpp and the GraphQL AI Working Group and co-host the German AI podcast Lethonium. Away from the keyboard: running, chess and piano.
Selected projects
Qwen3 GraphQL Embedding Model
Beats OpenAI text-embedding-3-large · 10K+ downloads
I fine-tuned Qwen3-Embedding-0.6B into a small embedding model that helps AI agents find their way around large GraphQL APIs. Ask a question in plain English and it points to the fields that answer it, so the agent does not need the whole schema in its context.
On the GraphQL AI Working Group benchmark it finds more of the right fields than OpenAI text-embedding-3-large (63.6% vs 59.8%), while being small enough to run locally on a CPU.
- Hugging Face
- Embeddings
- GraphQL
Worldbummlr
4M+ attractions · live product
AI audio guide: a geolocation-aware RAG platform that narrates stories about the places around you. Built the backend, ranking, route generation and data pipelines.
GraphQL MCP Server
Rust · schema discovery for agents
MCP server that lets agents explore large GraphQL APIs through dynamic type discovery, type dependency trees and optimized discovery pipelines.
Diffusion LLM
Text-to-SQL by text diffusion
A text-diffusion language model for Text2SQL, with training code and a playground that shows the query forming step by step.
yDE – Your Data Expert
1st place · Tech Europe Berlin AI Hackathon 2025
Voice-driven AI data analyst: upload a CSV and talk to it. Streams live charts, spoken answers and transcripts, and runs longer tasks like stance annotation while updating Google Sheets in real time.
Earlier projects
Neuron surgery on LLMs
Autoencoder-based feature extraction from LLM activations to amplify topic-specific neurons (e.g. SQL generation) at inference time.
On-device image classification
Small CNN running fully on-device in a Flutter app via ONNX.
Real-time news stance analysis
Trained a stance-detection model and built a platform clustering topics and sentiment across news headlines and Twitter in real time.
Chess via EEG
Playing chess by thought with a consumer-grade EEG, decoding imagined start and end squares into moves.
Algometer
Raspberry Pi device measuring algae and oil pollution, flow rate, salinity and temperature in the North Sea, reporting via LoRaWAN.
Open source
Contributor to the GraphQL AI Working Group since October 2025 and to llama.cpp (100K+ GitHub stars).
- Aug 2026
Embedding benchmark: Qwen3 GraphQL fine-tune
Merged into GraphQL AI Working Group
Added my fine-tuned Qwen3 embedding model to the working group benchmark: 63.6% field recall@50 vs. 59.8% for OpenAI text-embedding-3-large.
- Jan 2026
Experimental GraphQL MCP server
Merged into GraphQL AI Working Group
MCP server exposing any GraphQL schema to agents through two tools: semantic type search over embedded Type→field pairs, and validated query execution.
- Jan 2026
Metal POOL_1D kernel
Merged into llama.cpp
Added max and average 1D pooling kernels to the Metal backend, so models using 1D pooling run fully GPU-accelerated on Apple silicon.
Talks & podcast
- Jun 2026
“Teaching LLMs to Understand GraphQL with Schema-Aware Embeddings”
Talk · GraphQL Day at apidays 2026 · Amsterdam
Why general-purpose embeddings struggle with GraphQL type systems, and how schema-aware embeddings improve schema search, field retrieval and natural-language-to-query for MCP tools and agents.
- Sep 2025
“LLMs + GraphQL + MCP: A Blueprint for Scalable AI Tooling”
Talk · GraphQLConf 2025 · Amsterdam
- Jun 2024
“Effects of pharmacokinetics on arteries with in-stent restenosis in a fluid-solid framework”
Talk · ECCOMAS Congress 2024 · Lisbon
Session on numerical methods for the vascular system; with A. Ranno, S. Nerzak, K. Manjunatha and M. Behr (RWTH Aachen).
Lethonium
Podcast · in GermanDer 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 SpotifyResearch
Publications and thesis.
- 2025
A computational model of coronary arteries with in-stent restenosis coupling hemodynamics and pharmacokinetics with growth mechanics
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.
- 2024
Fluid–solid coupling solvers and physics-informed neural networks (PINNs) for arterial drug evolution simulation
Thore Koritzius · M.Sc. thesis, RWTH Aachen University
Coupled fluid–solid solvers and physics-informed neural networks to simulate how drugs evolve in arterial tissue.
- 2022
On-the-Fly Calculation of Model Factors for Multi-paradigm Applications
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.
Awards
- 2025
1st place, Tech Europe Berlin AI Hackathon
For yDE, a voice-driven real-time AI data analyst.
Konrad-Adenauer Scholarship
Government-funded scholarship for academic and extracurricular achievement.
1st prize, Jugend forscht
Computer Science, state level, Schleswig-Holstein.