Derick Lee

• Conducting research on LLM steering vectors to analyze semantic differences in social media data between two car brands

• Applied steering vectors extracted from 5,000+ tweets to Gemma 3B Instruct LLM from Hugging Face for brand-style generation

• Trained scikit-learn logistic regression model with 90% accuracy to classify brand tweets and evaluate accuracy on generated outputs

• Performed statistical and NLP analysis to generated text, achieving 95%+ accuracy and strong lexical similarity to 5,000+ tweets

• Engineering personalized AI tutor that learns what you teach and teaches what you learn. Backed by Antler VC and interviewed by ZFellows and Y Combinator

• Led implementation of a drag-and-drop class folder feature using Next.js, tRPC, and Prisma, including restructuring the database schema and designing new API routes

• Revamped the sidebar UI and chat interface using Tailwind CSS and TypeScript, delivering a modern, responsive experience to improve user navigation and engagement

• Maintained best practices in collaborative software development using Git, managing feature branches, code reviews, and safe production merges via pull requests

• Explored multi-agent reinforcement learning and Stackelberg-game framework, optimizing decision strategies for two-hop networks

• Implemented robust data-handling pipeline to export agent utility data in CSV format, reducing manual logging time by 80%

• Conducted scalability tests, analyzing network performance from 5 to 200 relay nodes and varying distances from source node

• Analyzed performance with NumPy, Pandas, and Matplotlib, identifying 50% utility dropoffs in deep reinforcement learning models