ms cs, machine learning
I’m an incoming research engineer in the PAIR lab at Georgia Tech, where I will work on evaluation and benchmarking for robot foundation models. Most of my research before that was in NLP for low-resource languages. I’ve built speech recognition and translation models for Formosan languages, some spoken by only a few thousand people. Working on low-resource languages got me interested in how models can still learn and generalize when data is limited, which is something I’d like to explore more in robotics.
Advised by Dr. Emily Prud'hommeaux. Fine-tuned NLLB-200 across 14 Formosan languages using FormosanBank corpora; two-stage fine-tuning with triple supplementary languages (Amis, Atayal, Rukai) proved most effective, improving FOR→EN BLEU by 50–79% for the four lowest-resource languages.
REU with Dr. Prud'hommeaux. Fine-tuned Wav2Vec2/XLSR-53 for 13 languages, reducing WER by up to 68% using cross-linguistic similarity metrics (Token Overlap) for transfer learning.
NSF REU with Dr. Rakesh Verma's ReDASers team. Supported XLM-RoBERTa experiments on cross-lingual toxicity detection and built an 85,000+ example Romance-language dataset.
Prud'hommeaux, E., Scheppat, H., Hernandez, N., & Le Ferrand, E.
Real-time OpenCV vision pipeline (HSV masking, contour/circularity analysis) tracking a Bluetooth-controlled 2-player robot car match. Pygame scoreboard; Claude API for live commentary.
view repo →
Developed the tree-search component of a hybrid tree-search + PPO agent for the MicroRTS real-time strategy environment. The combined agent achieved a TrueSkill rating of 33.7, outperforming the CoacAI baseline at 33.4.
view repo →
Predicting song popularity from lyrics across multiple classifiers: Naive Bayes, Decision Tree, KNN, SVM, Logistic Regression, DistilBERT.
view repo →
Z3 logical constraint solving to solve any configuration of the NYT "Pips" domino puzzle, with a Pygame interface to visualize and solve boards.
view repo →
Poster mapping the geographic density of U.S. data centers against regional utility costs.