Biography
I am a master’s student and Jefferson Scholar at the University of Virginia. I conduct research in Natural Language Processing (NLP) through the Information and Language Processing (ILP) Lab and Human-AI Interaction through the Ultimate Interface Lab. I also am a researcher in the Computational Approaches to Human Learning Research (CAHLR) Lab at UC Berkeley where I work on problems at the intersection of AI and Education.
- Natural Language Processing
- AI-assisted Decision Making
- Full-stack Development
Masters of Computer Science, 2024
University of Virginia
BA in Computer Science; BA in Financial Economics, 2023
University of Virginia
Experience
- Spearhead cross-campus development of NLP course-recommendation system (askoski.berkeley.edu)
- Direct back-end development and sprint planning for 14 engineer team
- (NeurIPS'23 GAIED) Designed a NLP feature: conversational recommendations for major selection advice
- Led mechanistic interpretability research with Distill-GPT-2 (42m param.) and LLM evaluation with GPT-4
- Generated PyTorch model code based on several papers without reproduction repositories
- Tuned ~10 hyperparameters to optimize the model training objective and final interpretability results
- Diligenced 4 B2B SaaS companies, interviewing 50+ industry experts, and modelling multi-geo. markets
- Leveraged SOTA AI models to improve efficiency of standard tasks, e.g. data classification, text summarization
- Automated several common tasks which saved the team ~5 hours per week of working time
- Developed COBA VowpalWabbit/coba for Python to evaluate contextual bandit (CB) algorithms
- Coded an online benchmark to train, evaluate, and analyze contextual bandit systems for research
- Implemented 3 CB algorithms, 2 data sources, and enhanced unit testing and parallel processing
Projects
Extending “Towards Monosemanticity”
Mechanistic interpretability by applying dictionary learning to Distill-GPT-2
Reel Recommendations
Leverage Reinforcement Learning and Language Models to perform movie recommendations for LetterBoxd users.
Corpus2Learn4Children
A toy educational app using language models to convert a document to slides for a 10-year old to learn from.
GenAI Risks and Benefits Blog
Blog proceedings of UVA’s Fall 2023 Seminar ‘Risks (and Benefits) of Generative AI and Large Language Models’
AskOski
The AskOski Project seeks to improve equity and achievement in higher education through human-centered AI research.
Coba
Coba is a powerful framework built to facilitate research with online contextual bandit (CB) algorithms.