GPU spec and price comparison site with 1,200+ daily unique visitors. I built its automated data collection pipelines, an LLM localization workflow (a frontier model writes the content, a local model translates it, automated checks reject bad output) and a scheduled LLM agent (LangGraph, Claude API) that checks scraped listings are filed under the right GPU.

Doctoral research applying deep reinforcement learning to multi-echelon inventory optimization; built a custom simulation environment and trained TD3, A2C and DQN agents (Stable Baselines3, PyTorch) to make inventory decisions. See my ▶️presentation for IVADO Digital October.

Competed solo in Kore 2022, a Google-sponsored simulation competition with strict per-turn compute and memory limits, and placed 4th of 469 (solo gold medal, $3,000 prize). Gained an edge by rewriting the simulation logic as vectorized NumPy matrix operations. A 📝write-up is on Kaggle.

1st place in Team Challenge #3 (Gunshot Audio Detection) with a custom CNN, and 2nd place among 100+ participants in the individual machine learning quiz.

Crowd-sourced Chinese-slang reference (“Urban Dictionary for Chinese”); used open-source LLM embeddings for similarity search to surface related terms.