Projects
RLHF for Open-Ended Instruction Following
An implementation of online policy-gradient RLHF (GRPO, DrGRPO, GSPO) for fine-tuning a 1.5B-parameter LLM on human preference pairs. The interesting parts include a replay-buffer method that builds preference pairs straight from policy rollouts, and a pessimistic reward-model ensemble that keeps the policy from gaming the reward.
Python · PyTorch · Modal · W&B
Python Compiler
A compiler I built from scratch in Java for a Python dialect. Covers the whole journey from raw source code through lexing, parsing, and type checking down to RISC-V assembly, including a simple web IDE that lets you watch each stage happen.
Java · Python · RISC-V · JFlex · CUP
Code Compass
A VS Code extension that ties together the pieces you juggle while debugging - the docs, the ticket, and the code actually involved - so the context you need lives in one place.
TypeScript · Python · Flask · PyTorch · OpenCV
Police Records Classification
A tool for making sense of large piles of police records, using few-shot learning to sort documents into meaningful categories while keeping the results interpretable enough to trust.
Python · SetFit · Sentence Transformers (SBERT) · Scikit-learn · Hugging Face · Plotly
Coding Task Extractor
A research helper that reads through papers and pulls out the coding tasks buried in their user studies, turning a tedious literature review into something an LLM pipeline can take the first pass on.
Python · OpenAI API · LangChain · Vector Stores and Search (FAISS) · Pydantic · Pandas
Procedurally Generated Escape Game
A 2D game engine that builds a fresh, fully solvable world every time you play, generating its rooms and corridors procedurally, and remembering your progress so you can save and replay.
Java · Algorithms · StdDraw / Algs4 · Git
Board Wizard
A companion for board-game night: ask a rules question in plain English and it answers straight from the rulebook, so nobody has to dig through the manual mid-game. Built on a RAG pipeline using the OpenAI API.
Python · Streamlit · LangChain · OpenAI API
Bear Territory
A mobile app that helps students find a good place to study nearby.
React Native · React · Expo · HTML · Navigation & Maps APIs
Roots
A design and prototyping project: a concept for a local food marketplace that nudges people toward greener choices, discovering nearby vendors, seeing the carbon footprint of a purchase, and earning small rewards for sustainable habits. Taken from research through to an interactive Figma prototype.
Figma · Product Design · Prototyping