This thesis presents a controlled empirical study evaluating the transformation of decoder-only LLMs (Sheared LLaMA) into powerful text encoders using the LLM2Vec method. The research compares frozen and fine-tuned configurations against a strong encoder-only baseline (DeBERTa) on NER and POS tagging tasks. The work involved extending the LLM2Vec codebase for HPC training and implementing MaChAmp’s gradual unfreezing strategy, effectively mitigating catastrophic forgetting and optimizing performance on domain-shifted datasets.
Large Scale Data Analysis
This project investigates how authenticity language is used across Yelp restaurant reviews, with a focus on potential cultural and racial biases in its application. Inspired by a 2019 study, the analysis explores whether words like "authentic" are used to signal different characteristics depending on cuisine type (e.g., “cheap” or “dirty” for Chinese/Mexican vs. “quality” for Italian/French). The work involves large-scale data processing with PySpark, influencer review detection, regional comparison, and predictive modeling using linear and logistic regression.
CopenhagenBuzz is an Android app that lets users discover, create, and favorite local events around Copenhagen, featuring Google Maps integration, photo uploads, and real-time updates using Firebase. Built as part of the Mobile App Development course at IT University of Copenhagen.
A collaborative MLOps project transforming a Python monolith into a modular, automated pipeline using GitHub Actions and Dagger, with model training, validation, and artifact management. Part of the Software Development & Engineering course at ITU.
A university project exploring privacy-preserving data publishing. This work implements and analyzes techniques like k-anonymity, differential privacy, and synthetic data generation to assess and mitigate re-identification risks in sensitive datasets.
This project applies machine learning techniques to classify greyscale images of clothing into five categories, using custom feature extraction, dimensionality reduction (PCA and LDA), and multiple classifiers including manually implemented Naive Bayes and neural networks. The work demonstrates and compares the effectiveness of these methods on a 15,000-image dataset from Zalando.
Natural Language Processing
This project focused on the comparison of DistilBERT with a custom fine-tuned LSTM on sentiment classification tasks using methods from Checklist. Analyzed model robustness to negation, paraphrasing, and named entity changes, revealing gaps in generalization and syntactic sensitivity.