APPLIED AI · RESEARCH & ENGINEERING应用 AI · 研究与工程实践
Model tuning, AI agents, and products模型调优、智能体与 AI 产品
.
I am an LLM engineer based in Guangzhou, working on language and multimodal model tuning, AI agents, and inference deployment. My work spans model training, CUDA operator development, and product delivery, alongside independent AI products developed in my spare time.我是杨瑞卿,现居广州,专注于大语言模型与多模态模型调优、AI Agent 及推理部署。我的实践覆盖模型训练、CUDA 算子开发和产品交付,也在业余时间开发个人 AI 产品。
2BXcoder tokens / dayXcoder 日均 Token
2 billion tokens/day including cached tokens; average daily platform token usage.日均 20 亿 Token(含缓存 Token),口径为平台 Token 用量。
AAAI2025 · Second-author paper2025 · 第二作者论文
Ship license plate recognition research · CCF A conference.复杂场景船牌文本识别研究 · CCF A 类会议。
10+Customers served服务客户
Customer delivery at Alibaba Cloud, mainly for global small and medium-sized businesses.阿里云客户交付工作,主要服务全球中小企业。
01 /
What I work on研究与实践方向
01Language & multimodal model tuning语言与多模态模型调优
↗ 02AI agents & workflow automationAI Agent 与工作流自动化
↗ 03Inference deployment & CUDA operators推理部署与 CUDA 算子
↗ 04Scene text recognition自然场景文本识别
↗
02 /
Independent projects个人产品
Built in my spare time业余时间的产品实践
CREATE · GENERATE · REFINE创作 · 生成 · 编辑
AI characters and video creationAI 角色与视频创作平台
An independent product developed in my spare time, combining image generation, video generation, and an infinite canvas. All platform models are deployed on local GPUs, with LoRA training and ComfyUI workflow tuning for specific creative scenarios.业余时间开发的个人产品,整合图像生成、视频生成与无限画布。平台模型均基于本地 GPU 部署,面向特定创作场景开展 LoRA 训练,并基于 ComfyUI 进行工作流编排与定向调优。
- LoRA
- ComfyUI
- Local GPU
- Multimodal Generation
ic.xshow.live↗
CODE · BUILD · SHIP编程 · 构建 · 交付
AI coding model servicesAI 编程模型服务平台
An independent product developed and launched in my spare time, providing coding model services through an OpenAI-compatible API and tiered subscriptions. Average daily platform usage is 2 billion tokens/day including cached tokens.业余时间创建并上线的个人产品,通过兼容 OpenAI 的 API 和分层订阅提供 Coding 大模型服务。平台日均 Token 用量达 20 亿(含缓存 Token)。
- Coding Models
- OpenAI-compatible API
- Streaming
- Tool Calling
xcodex.vip↗
03 /
Selected research代表研究
AAAI 2025Second author第二作者
Research on ship license plate recognition in challenging real-world scenes, with a dataset of 36,000 real-scene images and a recognition method drawing on CLIP and MAE. The project reports 84% recognition accuracy, compared with 70% for GPT-4o in the same project evaluation; these figures apply to this task and evaluation setting. Related patent publication: CN119251820A.面向复杂真实场景的船牌文本识别研究,发布包含 3.6 万张真实场景图像的数据集,并基于 CLIP 与 MAE 思想构建识别方法。该项目报告的识别准确率为 84%,GPT-4o 在该项目对照评测中为 70%;这些数字仅对应这一任务与评测设置。相关专利公开文献:CN119251820A。
2025.02 — Present / 至今
Alibaba Cloud, Alibaba Group阿里巴巴 · 阿里云智能集团
Algorithm SA-TL / LLM Algorithm Engineer · Alibaba Cloud International阿里云国际|算法 SA-TL / 大模型 AI 算法工程师
Tune Qwen3 models, including Qwen3.8-27B Coding enhancements to support Qoder and Qwen3.6-35B-A3B optimization for the Smart Green Channel (智慧绿通) use case.负责 Qwen3 系列调优,开展 Qwen3.8-27B Coding 能力增强,为 Qoder 提供编程模型支持;面向智慧绿通场景优化 Qwen3.6-35B-A3B。
2024.12 — 2025.02
Shumei Technology (Beijing)北京数美时代科技有限公司
Algorithm Engineer · Tianjing Vision Team算法工程师|天净视觉组
Serve as technical lead for OCR and object detection in content moderation, including politically sensitive content, sexual content, and promotional redirects; handle data selection and cleaning, model iteration, architecture updates, and difficult customer cases.担任 OCR 技术负责人及物品检测负责人,面向涉政、色情、广告导流等内容审核场景,负责数据筛选清洗、模型迭代与架构更新,结合客户需求解决疑难样本。
2024.03 — 2024.11
Brain and Intelligence Lab, Tsinghua University清华大学脑与智能实验室
Algorithm Engineer算法工程师
Conduct full-parameter training and LoRA fine-tuning of Llama, Qwen, and Mistral with C-RLFT on eight H800 GPUs, and evaluate models with OpenCompass; develop custom JAX / CUDA operators for BF16 sparse matrix multiplication and spiking neural network training.基于 C-RLFT 在 8 卡 H800 上开展 Llama、Qwen、Mistral 全参数训练与 LoRA 微调,使用 OpenCompass 评测;编写 JAX / CUDA 自定义算子,支持 BF16 稀疏矩阵乘法及脉冲神经网络训练。
2023.08 — 2024.02
Mech-Mind Robotics (Beijing)梅卡曼德(北京)机器人科技有限公司
Deep Learning Computer Vision Intern深度学习视觉算法实习生
Develop the ML-DLK deep learning C++ SDK and a PyTorch → ONNX → ONNX Runtime deployment pipeline; implement custom CUDA operators such as DeformConv2D and deploy 3D point-cloud algorithms and SegGPT.研发 ML-DLK 深度学习算法 C++ SDK,完成 PyTorch → ONNX → ONNX Runtime 部署,编写 DeformConv2D 等 CUDA 算子,并部署 3D 点云算法与 SegGPT。
Zhejiang Gongshang University浙江工商大学
Academic Master's Degree in Computer Science and Technology计算机科学与技术 · 学术硕士
Top 5% in the program's comprehensive ranking; research focused on OCR in complex natural scenes.专业综合排名前 5%;研究方向为复杂自然场景 OCR。
2021.09 — 2024.06
North China University of Science and Technology华北理工大学
Bachelor's Degree in Communication Engineering通信工程 · 本科
Top 5% in the program's comprehensive ranking; passed CET-4 and CET-6.专业综合排名前 5%;通过 CET-4 / CET-6。
2017.09 — 2021.06