Projects

• Currently working with professor Kay Giesecke to understand the impact of natural disasters on mortgage backed securities. 

In this project, I worked with Prof. Peter Washington from the University of Hawaii. We developed a nove machine learning approach that uses only non-invasive predictors in order to predict patients' blood glucose levels. The idea of the project was to lay the foundations for the potential replacement of  CGM devices through self-supervised, individualized modelling.

• Github can be found here

• Preprint can be found on medizinische medRxiv

In this project, my teammates and I developed a basketball betting product that is analogous to Bermuda Options and priced it using reinforcement learning. Using features extracted from basketball-reference.com, we applied least squares policy iteration (LSPI) and Deep Q Learning (DQL) to price bets on a players points scored in a game with cash-out offered at the end of every quarter. 

• Github can be found here

• Final presentation can be downloaded here

In this project, we used a basic implementation of BERT to test the effectiveness of Low Rank Adaptations (LoRA) for fine-tuning large language model in a multitask setting. The goal was to see whether LoRA would be able to produce embeddings that perform simultaneously well on sentiment analysis, paraphrase detection and semantic textual similarity scoring. We built a custom implementation of LoRA for a multitask setting and ran experiments on the three tasks in comparison with various BERT baselines. We found that LoRA performs comparably to conventional fine-tuning and were able to reproduce the findings of the original authors. This project was the final project for CS 224N Natural Language Processing with Deep Learning. Both Marco and I received an A+ in the class. Due to course policy, I am unable to post our project publicly, please email me if interested. 

In this project, my teammates and I developed a new data imputation technique based on Kotelnikov et al.'s TabDDPM (2022), a denoising diffusion probabilistic model for tabular data. Rather than modelling whole data distributions, we modified the model to instead apply it to the problem of data imputation. Our intial results were promising, particularly on categorical variables, where we found that TabDDPM outperforms conventional data imputation methods. This project was the final project for CS 229 Machine Learning and received the top grade, scoring 100/100 points. 

• Github can be found here

• Report can be downloaded here

• Poster can be downloaded here

In this thesis project, I analysed the development of the sentiment of Twitter users towards alternative fuel vehicles over the past five years (2017 - 2022). I collected 5.000.000 Tweets on electric vehicles and synthetic fuels using tweepy and the Twitter API.  Using transformers library from HuggingFace, I performed sentiment analysis on all tweets I then created an algorithm based on relative word frequency to determine which events were trending and driving sentiment in particular months. Mapping these events to a the development of the sentiment in tweets observed, I was able to analyse the impact of certain events on public attitudes towards electric vehicles.  

• Full thesis pdf can be downloaded here

• Github Repository can be found here

• Won the SAP data analytics challenge of the Technical University of Munich's official hackathon. 

• Developed a product that automatically creates TikTok-style videos to aid management decision-making based on sales data.

• Held final winning presentation

• Project overview can be found here

• Github repository can be found here