AI AGENTS / MACHINE LEARNING
My work mainly involves building applications with AI agents.
I’m also learning about AI infrastructure.
Core contributor to RL-Kernel ↗
I work with LLMs on practical tasks and have a background in machine learning and deep learning, including how models work and how they are structured. I use Python, C++, and C.
I’m also learning about reinforcement learning and following projects such as vLLM. Along the way, I contribute to open-source projects in this area.
I contribute to infrastructure for reinforcement learning (RL) post-training, working on logprob operators, numerical comparisons between training and inference, and ROCm support.
Contributions as of 2026-09-11 · Upstream RL-Align/RL-Kernel ↗
Do LLMs learn rules or memorize patterns? This project attempts to reproduce a study of LLM arithmetic, using numerical addition, symbol mappings, and commutativity tests to examine arithmetic capabilities and generalization.
Dataset specifications are from the repository README. Code and datasets were developed with LLM assistance.
Code & experimental setup ↗JOURNAL ARTICLE · FIRST AUTHOR
基于BERT-CLS-ATT模型的虚拟主播评论情感分类算法
English title translated from the Chinese original.
Combines BERT pretrained on domain-specific data with an attention mechanism to classify sentiment in comments on virtual streamers, followed by TF-IDF keyword extraction.
On the virtual streamer comment dataset: F1 > 83%Accuracy > 84%
The study proposes BERT-CLS-ATT, combining BERT pretrained on additional domain-specific data with an attention mechanism to classify sentiment in comments on virtual streamers. Classification results and TF-IDF keywords support content analysis. Further experiments on public datasets explore the architecture’s generalizability to sentiment classification.
PATENT · GRANTED 2021 · CO-INVENTOR
基于离散分数变换和混沌函数的图像加密方法及装置
English title translated from the Chinese original.
Image encryption combining chaotic functions, discrete fractional transforms, and data scrambling. CN110414250B ↗