Android-first pantry and meal management MVP with FastAPI, SQLite, Expo React Native, local vision-based ingredient recognition, and read-only recipe recommendations.

A local-first Windows desktop app with an aquarium-style token tank UI for tracking AI coding sessions and estimating remaining quota.

Hybrid planning agent for decision-making under partial observability (PG-WMA)

From-scratch GPT-style Transformer that achieves 99%+ accuracy on 4-digit arithmetic via autoregressive scheduled sampling.

A production-style shadow trading framework, separates the alpha, risk, and execution layers.

AntWorld – Multi-ant grid simulation with local memory, pheromones, and 5×5 communication（多螞蟻網格模擬，用於世界模型與強化學習實驗）

A compact, fully self-contained Transformer framework for LoRA fine-tuning and text generation.

STaR × S1 math pipeline on Qwen2.5-1.5B. LoRA, strict Final: format, ~20–30% acc (OpenR1-Math split).

Fine-tuning Phi-2 with LoRA for grid-based spatial reasoning and Chain-of-Thought (CoT) inference.

A custom POMDP maze environment for studying agent reasoning, uncertainty, partial observation, and world-model learning. Produced structured trajectories for training GPT-based reasoning models.