Data Science & Big Data Technology @ University of Jinan (2024–2028) · Intern @ Institute of Automation, Chinese Academy of Sciences (CASIA).
I build at the intersection of recommendation systems, user growth, and AI agents — from research to production. I've contributed merged PRs to projects with 190K+ stars (n8n) and 36K+ stars (LightRAG).
Python C/C++ TypeScript PyTorch Transformers Scikit-learn XGBoost LangChain FAISS FastAPI
| Project | ✨ / Stars | Contribution |
|---|---|---|
| n8n | 193K | Fixed v2.24.0 regression crash in Form Trigger node. Merged. |
| LightRAG | 36.6K | Refactored NDJSON streaming parser, fixed stream error bug. Merged. |
| TensorTrade | 6.3K | Added short-selling support to OMS & Action Schemes. Submitted. |
| AI_HR_Project | — | Batch resume screening pipeline with LLM scoring. Merged. |
| WeMD | — | Removed hardcoded font-size in table renderer. Merged. |
| Project | Highlights | Tech Stack | Repo |
|---|---|---|---|
| OneTrans & MixFormer: Unified Transformer Recommender | Tencent Ads Algorithm × KDD Cup — End-to-end single-stage Transformer unifying sequence modeling & feature interaction, validated on million-scale data. | Transformer | Private (NDA) |
Game Churn Prediction & Persona Platform 2025.3–4 |
National E-commerce "Sanchuang" Competition — XGBoost churn + KMeans clustering + BERT sentiment analysis on 46K reviews. | XGBoost, KMeans, BERT | Repo |
| Claude MD Editor — AI-Powered Desktop Markdown Editor | Electron + React + TypeScript app with CodeMirror 6, KaTeX, Mermaid, and embedded Claude Code AI terminal. One-click multi-format export. | Electron, React, TypeScript, CodeMirror 6, KaTeX, Mermaid | Repo |
| arXiv MCP Server with DeepSeek Translation | MCP-protocol academic search: Chinese→English translation, arXiv retrieval, and semantic re-ranking via FastAPI + fastapi-mcp SSE. | FastAPI, fastapi-mcp SSE | Repo |
| AI Customer Service with LoRA + RAG + LLM | Three-tier architecture: RAG (Sentence-Transformers + FAISS) → DeepSeek candidate generation → LoRA-tuned MacBERT for reply ranking. | RAG, FAISS, DeepSeek, LoRA, MacBERT | Repo |
| Cat AI System for "Whispers of the Heart" | Tencent Game Creation Competition — Three-tier cat AI: RL strategy layer + custom behavior tree engine + local DeepSeek-R1-Distill-Qwen-1.5B for emotional monologues. | RL, Behavior Tree, DeepSeek-R1-Distill-Qwen-1.5B | Repo |
I enjoy building systems rather than isolated demos.
Over the past few years, most of my work has naturally converged around three topics: recommendation systems, user growth, and AI agents. Though they seem different at first glance, they all revolve around the same core question: How can intelligent systems make better decisions with limited information?
That question has driven me to train recommender models, design analytics platforms, deploy RAG applications, and explore multi-agent workflows.
Outside of my own projects, I'm passionate about open source. For me, every merged PR is more than code — it's a window into how world-class engineering teams design, review, and maintain software at scale.
I thrive on building in public, sharing what I discover, and turning ambiguous ideas into working prototypes. I believe great software comes from continuous iteration, honest feedback, and restless curiosity.
If you're interested in recommender systems, AI agents, open source, or just want to riff on interesting technical ideas — I'd love to connect!
📖 中文版本 (点击展开)
我是济南大学数据科学与大数据技术专业本科生（2024–2028），目前在中国科学院自动化所实习。
我的方向聚焦于推荐系统、用户增长与 AI Agent，工作覆盖研究到工程。我曾向 19 万 Star 级项目（n8n）和 3.6 万 Star 级项目（LightRAG）提交过被合并的 PR。
Python C/C++ TypeScript PyTorch Transformers Scikit-learn XGBoost LangChain FAISS FastAPI
| 项目 | ✨ / Stars | 贡献 |
|---|---|---|
| n8n | 193K | 修复 Form Trigger 节点在 v2.24.0 的回归崩溃。已合并 |
| LightRAG | 36.6K | 重构 NDJSON 流式解析器，修复流错误处理 Bug。已合并 |
| TensorTrade | 6.3K | 为 OMS 和 Action Schemes 增加做空支持。已提交 |
| AI_HR_Project | — | 批量简历筛选流水线，基于 LLM 多维评分。已合并 |
| WeMD | — | 移除表格渲染器中硬编码的字号。已合并 |
| 项目 | 简介 | 方法 / 技术栈 | 仓库 |
|---|---|---|---|
| OneTrans & MixFormer：统一 Transformer 推荐框架 | 腾讯广告算法大赛 × KDD Cup — 将序列建模与特征交互整合为端到端单阶段 Transformer，在百万级交互数据上验证。 | Transformer | Private (NDA) |
游戏流失预测与用户画像平台 2025.3–4 |
全国大学生电子商务"三创赛" — XGBoost 流失预测 + KMeans 用户分群 + BERT 评论情感分析（46K 条）。 | XGBoost, KMeans, BERT | Repo |
| Claude MD Editor：AI 桌面 Markdown 编辑器 | Electron + React + TypeScript 桌面应用，集成 CodeMirror 6、KaTeX、Mermaid，内嵌 Claude Code AI 终端，一键多格式导出。 | Electron, React, TypeScript, CodeMirror 6, KaTeX, Mermaid | Repo |
| arXiv MCP 学术检索服务（接入 DeepSeek） | 基于 MCP 协议的学术检索：中文翻译→arXiv 查询→语义重排序，提供 FastAPI + fastapi-mcp SSE 端点。 | FastAPI, fastapi-mcp SSE | Repo |
| 智能客服：RAG + LoRA + LLM 三层架构 | RAG（Sentence-Transformers + FAISS）检索 → DeepSeek 候选生成 → LoRA 微调的 MacBERT 进行回复重排序。 | RAG, FAISS, DeepSeek, LoRA, MacBERT | Repo |
| 《猫语心声》游戏 AI 决策与语言系统 | 腾讯游戏创作大赛 — 三层猫咪 AI：RL 策略层 + 自研行为树引擎 + 本地 DeepSeek-R1-Distill-Qwen-1.5B 情感独白生成。 | RL, Behavior Tree, DeepSeek-R1-Distill-Qwen-1.5B | Repo |
我喜欢做完整的系统，而不是孤立的 Demo。
过去几年，我的工作自然收敛到三个方向：推荐系统、用户增长和 AI Agent。表面不同，内核一致——智能系统如何在信息有限时做出更好的决策？
这个问题驱动我训练推荐模型、设计分析平台、部署 RAG 应用，也探索多 Agent 协作。
开源对我来说不止是代码贡献，更是向世界级工程团队学习设计、审查和可维护性思维的绝佳窗口。
我习惯“公开学习”，边做边分享，把模糊的想法快速变成能跑的原型。我相信好软件来自持续迭代、坦诚反馈和永不满足的好奇心。
如果你也对推荐系统、AI Agent 或开源感兴趣，或者只是想聊聊有意思的技术脑洞，欢迎联系我！
不诱于誉，不恐于诽，率道而行，端然正己。