Computer Science Student · Tübingen
Armin Burkhardt
I build at the intersection of quantitative finance, machine learning and interdisciplinary applications.
Scroll
About
I study computer science in Tübingen. What interests me most is what happens when modern machine learning models reach complex domains where mistakes are expensive - trading, law, medicine, etc. Alongside that I am building the Tübingen Quant Society, a student initiative meant to close exactly that gap between theory and practice.
- Studies
- B.Sc. Computer Science, University of Tübingen, 2024 - 2027
- Role
- Student Research Assistant, Institute for Bioinformatics and Medical Informatics (IBMI) Tübingen
- Focus
- Quant Finance · Machine Learning · Interdisciplinary ML
- Initiative
- Tübingen Quant Society, Co-Founder
Projects
Co-founded in Tübingen to give students hands-on access to quantitative finance and organize events. For the initiative I co-developed the bilingual website: Next.js App Router, statically served, with a SAML 2.0 service provider for university login.
- Next.js
- TypeScript
- SAML 2.0
- Vercel
Built at the HackTheLaw Hackathon at the University of Cambridge for Legora's "The Sparring Room" challenge. Rather than answering questions, the system attacks the user's reasoning and forces them to defend their position - legal sparring instead of a research assistant. Python backend on FastAPI, frontend in React and Vite.
- Python
- FastAPI
- React
- Vite
- LLMs
An experimental model on a dense Gemma4-style backbone: a single MoE block is applied for several iterations (LoopLM-style), rerouting tokens on every pass. Heterogeneous experts - self-attention, cross-attention, retrieval and MLP - share one router. Multi-token prediction as an auxiliary objective, trained in FP8 and NVFP4 among others.
- PyTorch
- Transformer Engine
- CUDA
- MoE
- MTP
Currently training - the repository goes public once the run completes.
Everything that would normally need a glance at the screen - weather, directions, calendar, messages, alarms, device settings, the camera as a description of your surroundings - happens through a Voice to Voice Interface. An LLM (optionally local) decides which of the on-device tools to call and acts on the user's behalf; replies are shaped for German text-to-speech and tailored to the target audience.
- Android
- Kotlin
- LLM Tool-Use
- TTS
- ASR
- RAG
University team project - the repository is not public at the moment.