Projects
Stuff I've built to learn and apply said learning to my passions.
FaceGuard—a MacOS facial recognition security system
Never worry about your Mac again. This lightweight facial recognition software detects anyone invading your privacy—one email away. Without needing GPUs, it's quickly trained on your face and will send you a video and pictures of intruders. Fully open-source.
SmartSnake — a self-training PyTorch Snake agent
Snake, but actually smart—I built a self-training PyTorch agent that learns in real time. A compact 2-layer net (11→256→3) reads hazards, heading, and food direction to choose left/right/straight in a Pygame loop; experience replay (100K/1K), TD(0), and ε-decay keep it stable and sample-efficient. With best-model checkpointing and live/mean score plots, it jumped 733% by game 100 and leveled off around game 500.
myHomer — your agentic Columbia advisor
Columbia, decoded — a campus-aware RAG chatbot that searches a per-user ChromaDB index of syllabi, policies, and resources to return grounded answers with citations. Task-tuned prompts, query rewriting, multi-turn memory, and safety/retrieval guardrails keep context on target; an agentic add-on auto-registers for classes the instant seats open.
cpyAnalyzer — an ML code analyzer
Pickle-based models contain high cybersecurity risks—so I built this analyzer to mitigate them. By drawing class-hierarchy dependency graphs, cpyAnalyzer allows you to ensure important objects aren't transformed into malicious ones capable of taking over your system.
