Madrid, Community of Madrid, Spain
My name is Marcel, born in Germany, currently studying Deep Learning in Madrid. I am a young man, driven by curiosity. Since I was a child, I have been interested in technology. I am fascinated by looking at what clever ideas humankind comes up with and I would love to be able to create something special myself that brings us forward! During my dual studies I discovered my affinity for coding and artificial intelligence. I love the idea of utilizing artificial intelligence to give back people their precious time to focus on things that make us humans truly special: Unleashing creativity, solve complex problems and interact with each other. So far I am trying to learn as much as I can to pave the way for a great career!
Continued working on traffic scenario generation project: - optimization of earlier created scripts - creation of custom tokenizers - trained Sequence-to-Sequence-Models as finetuning alternative Ended with capable Seq2Seq-Models that are noticeably smaller than the finetuned ones. Pioneering work for Q&A bot: - data exploration and cleaning - reformatting data to use it for training using other LLMs - perform training and finetuning - implement evaluation methods to compare different models with each other
Self-driven research project about finetuning Large Language Models to generate XML-like file structures used for traffic scenario generation. This included, for instance: - research / learning about finetuning techniques, main focus on LoRA - select usable models based on it's characteristics - Data Exploration and adaptations of XML-like data - create robust scripts for finetuning, inference, dataset generation etc. - setup multi-GPU usage including reimplementation of Distributed Data Parallelism - usage of Python Frameworks such as PyTorch, Huggingface Transformers, Unsloth, Pandas, PyPlot, Hydra - UI creation using Streamlit - Prompt Engineering for data manipulation using other LLMs - comparison of capable finetuned models - presenting results to the whole team in a presentation At the end, I had capable models to do the scenario generation based on Qwen2.5 and Llama3.
I learned and applied the fundamentals of software development in a Web3 and SSI team. Later, I taught myself necessary knowledge in Machine Learning, then switched into an AI team. There, I worked on a project focusing on predicting customer's intents with neural networks. For instance, this included: -coding data processing or training scripts -creation and adaptation of Machine Learning Pipelines -providing model evaluation metrics -usage of Git and GitLab -working in a Scrum-based team environment I dedicated my bachelors thesis to this project. As part of it, I trained and evaluated the performance of various Transformer architectures for this intent prediction use case.