Fine-tuned Vision Transformer (ViT-B/16) on 27K EuroSAT satellite images across 10 land-use classes. Compared DINO self-supervised vs. supervised pretraining under linear, partial, and full fine-tuning. Includes attention map visualization and interactive results dashboard.

Prompt-engineered GPT-4 and Qwen 2.5 through RESTful API to create an AI agent tutoring non-native English speakers. Used spaCy NLP to achieve multi-style agent switching with 95%+ accuracy.

Static analysis tool that maps Python codebases via AST parsing, cross-file import resolution, and function-level call graphs. Features backward slicing that reduces code reading by 50%+, plus interactive D3.js visualization.

Engineered 10+ features from census data and implemented ML classifiers from scikit-learn. Optimized via GridSearchCV achieving 92% accuracy and 85% F-score for income-level classification.

Official website for NGO @ UC Berkeley. Used Bootstrap frontend with Notion as database. Attracted 50+ new membership applications per semester. Deployed via SSH on Berkeley's Open Computing Facility.

Web scraping bot using Selenium and Beautiful Soup to collect rental information across San Francisco. Automatically uploads structured data to Google Sheets via Sheety API.