Improved tool-calling accuracy 30 points over base via custom GRPO reward shaping, and raised accuracy a further 10 points by fine-tuning Qwen3.5 4B with LoRA/PEFT and Hugging Face TRL. Cut latency 50% and eliminated per-call API spend by replacing a frontier model with a locally served 4B SLM, and authored a preprint on progressive tool disclosure in tool-calling agents.

Devised a full-stack (Next.js + FastAPI) app providing LLM-powered Japan travel guidance and recommendations. Fine-tuned Granite 8B (via MLX) using 50K Japan specific question/answer pairs.

Built a RAG application (Next.js+Chroma DB) supporting natural language queries across 10+ operational documents with accurate source attributions.

Developed a Flask application to handle incoming customer SMS messages via Twilio API and deliver GPT-generated responses, cutting average customer wait time from two business days to near-instantaneous.

Trained an agent to complete Super Mario World 1-1 using reinforcement learning techniques while teaching peers the basics of RL.

Classified emotions of faces from downloaded images using computer vision techniques.

Led a 7-person team to build a Random Forest–based fare prediction model with an average error margin of $100. Automated a daily data pipeline using GitHub Actions and constructed an ETL pipeline for handling a 31 GB dataset.