Georgia Institute of Technology
MS in Computer Science
MSCS @ Georgia Tech
About me
I am a computer science graduate student who enjoys building reliable software systems and turning research ideas into working prototypes. My work spans full-stack engineering, applied machine learning, LLM systems, and AI agents.
I am especially interested in ML, LLMs, DRL, World models, and Embodied AI. Outside of research and engineering, I love snowboarding and teaching as a CASI snowboard instructor.
Education
Academic Background
University of British Columbia
BS in Computer Science
Research
Research Interests
- ML
- LLM
- DRL
- World Model
- Embodied AI
Projects
Software / Systems
AI onboarding assistant
Inductify
An AI-powered onboarding chatbot that lets new hires query 10K+ internal documents with cited answers under 5 seconds, backed by a RAG pipeline and retrieval evaluation.
Fault-tolerant processing system
Data Processing Web System
An async JSON data-processing platform designed around a FastAPI API, Redis-backed Celery workers, PostgreSQL task state, and containerized worker scaling for concurrent, failure-aware processing.
Full-stack web application
UniQuery
A web application that lets users query UBC course data efficiently. The project focuses on practical data ingestion, filtering, and a clean query experience for academic information.
Research
Research Projects
Independent research - LLM distillation for vision-language models
On-Policy Distillation for Vision-Language Reasoning
Transferred on-policy distillation (DAgger-style, per-token KL to a teacher on the student's own rollouts) to Qwen3-VL on ChartQA. A 2B student trained on 300 questions matches supervised distillation on 3,000 (+5.8 points at equal data, paired 95% CI excludes 0), and a token-level analysis shows the teacher's feedback lands on answers and format before chart values.
Publication - ACM GoodIT 2026 - Pisa, Italy
FireFair: Equity-Adjusted Multi-Agent Triage for Wildfire Ignition Forecasting
FireFair audits a California wildfire forecaster by the CDC/ATSDR Social Vulnerability Index, cuts the equal-opportunity gap by 75% at a 1.6pp F1 cost, and routes marginal signals into visible multi-agent verification.
Multi-agent reinforcement learning
Soccer-Twos Multi-Agent RL
A PPO self-play curriculum for 2v2 Soccer-Twos, using reward shaping, separated policy/value networks, and mixed opponents to reduce catastrophic forgetting and win a 72-team tournament.
Research project / preprint
GNNs for Product Recommendation
A graph learning study on Amazon co-purchase recommendation, comparing LightGCN, GraphSAGE, GAT, and PinSAGE under link prediction settings with practical trade-off analysis.
Contact