Zekai Huang

SELECTED PROJECTS

Built to learn, tested by use.

A reference implementation for LLM agent runtime engineering: native ReAct loop, sandbox isolation, flat-file memory, and SSE streaming.

Qwen3-8B Text2SQL post-training from supervised fine-tuning through reinforcement learning.

A Codex skill that turns photos and memories into craftable fuse-bead keepsakes with adaptive grids, real palettes, and structural-growth GIFs.

SELECTED WRITING

从 Agent Harness、Post-Training 和项目重构中学到的事：先跑起来不等于做对，AI 可以提方案，但判断的尺子必须在自己手里。

从 Self-Attention 出发，逐层拆解 Transformer 的编码器、缩放点积注意力、多头注意力、掩码与位置编码，配合公式、架构图和 PyTorch 实现。

从 One-Hot、词袋和 TF-IDF 到 Word Embedding、Tokenizer 与 BPE，系统理解文本如何被映射到连续语义空间，并附完整代码实现。