Worked on FedPMR, a personalized prototype-based federated learning framework for multi-center accelerated MRI reconstruction. The system supports cross-institution model collaboration while keeping raw medical data local.

Result: Validated on three public datasets and one private dataset from The First Affiliated Hospital, Zhejiang University. Communication cost was reduced from 236.70 Mb to 8 Kb, with stronger reconstruction performance on heterogeneous sites.

Authors: Jing Wang, Tong Wang, Siyuan Luo, Xiaoran Guo, Jiayi Yin, Jing Li, Zehua Yuan, Gang Yu, Bo Lin.

提出一种面向临床多源异构数据的冠心病多模态风险预警方法。针对心电信号（ECG）、超声心动图视频与结构化临床表格数据三类异构模态，设计独立的深度特征提取网络，并通过 InfoNCE 对比学习实现跨模态特征空间对齐。

Innovation：引入基于可学习标识符的模态丢弃策略解决真实临床中的整模态缺失问题，结合交叉注意力机制进行特征动态加权融合，最终通过深度生存分析输出个体化动态生存曲线，为临床分诊提供时间维度的量化依据。

Field：数字医疗 · 医学人工智能 · 多模态深度学习

一个面向 AI 生成图像策略评估的在线盲测平台，用于把“看图打分、方案对比、结论导出”这类评审流程产品化。

Built a lightweight web-based blind testing platform for evaluating AI-generated images across multiple strategies. Reviewers can score anonymous options, add quick labels and comments, then reveal strategy identities after review.

Product value: Turns subjective visual comparison into structured, exportable evaluation data, making model and strategy discussions easier to align across product, design, and technical teams.

一个把个人主页虚拟形象 Loopi 做成可评测、可迭代 AI 产品实验的开源系统：真实访客反馈进入数据库，Agent 读取反馈、评估红线、生成候选并给出 staging 建议。

Built a personal-brand AI companion system for yuanzehua.me, turning a homepage mascot into a measurable product loop with visitor feedback, version records, candidate prompts, evaluation reports, and human-gated release decisions.

Product value: Demonstrates how AI-native product work can connect hypothesis, measurement, controlled generation, evaluation, and deployment without letting novelty override homepage fit or personal-brand clarity.

一个面向个人网站长期维护的 Codex Skill，把“读项目、找源文件、控制实验、验证改动、谨慎发布”沉淀成可复用的 AI 协作流程。

Built a reusable Codex skill for maintaining personal websites as long-lived personal products. It separates inspection, maintenance, experimentation, publication, and rollback, helping Codex avoid editing generated output or publishing without explicit authorization.

Workflow value: Turns repeated homepage updates into a source-aware, verification-first operating system, making AI collaboration safer for content updates, visual experiments, production checks, and rollback planning.

Built a knowledge-base Q&A system over PDF, DOCX, and Markdown documents using OpenAI APIs.

Design: Combined semantic chunking, embeddings, vector retrieval, and prompt design to reduce omissions and hallucinations in complex document Q&A, turning raw materials into an interactive knowledge base.

Designed a FedAvg-based multi-client training framework for privacy-sensitive behavior recognition, enabling model collaboration without moving local data.

Balance: Focused on the tradeoff between model performance, data locality, privacy protection, and deployability.