Xin Yan

Xin Yan

Emotional Intelligence

About

I believe the real value of large language models lies not in answering everything, but in being trustworthy and useful in specific contexts.

I graduated from East China University of Science and Technology, majoring in Automation, and currently work as a Senior Algorithm Engineer at Cylingo Group. I lead the R&D of the Xinyuan series of domain-specific large language models. Over the past few years, I have focused on affective intelligence and domain-specific large model deployment, building full-cycle pipelines from data curation and continual pre-training to SFT, RLHF, model deployment, and community operations across scenarios including mental health, healthcare, retail, and community applications.

I led projects including Xinyuan-LLM, Xinyuan-VL, MindChat, Sunsimiao, ColugoMum, and OXiaoPeng. Across the open-source projects I have led and contributed to, my work has collectively earned 20,000+ GitHub Stars and appeared multiple times on GitHub Trending. I have also published 1 book on LLM application development, 4 papers, and hold 1 utility model patent and 3 software copyrights. My work has received 7 national-level awards and over 10 provincial/municipal awards.

The question I keep asking: How can we make AI more usable, safer, and more sustainable in real human scenarios?

News

  • 2026.05: Co-authored LangChain Large Model Application Development: From Beginner to Practice, Tsinghua University Press.
  • 2026.01: First Prize in the 2025 Synthetic Data Competition · Lingxi AI for Mental Health track.
  • 2025.09: Speaker at the 2025 Apsara Conference session “Doing AI, the Gen-Z Way”.
  • 2025.07: Bronze Award in the 2024 China International College Students’ Innovation Competition.
  • 2025.07: Joint interview by Alibaba Cloud Tongyi Lab and ModelScope community.

Experience

Cylingo Group | Senior Algorithm Engineer
March 2024 - Present

Problem: General-purpose LLMs tend to give generic responses in mental health and education scenarios. They lack domain knowledge and struggle with safety and empathy boundaries.

Solution: Led the full-cycle R&D of the Xinyuan series of domain-specific LLMs, covering de-identification, deduplication, quality scoring, continual pre-training, SFT, RLHF, and multi-dimensional evaluation across 45M+ data samples. Coordinated a team of 5-20 algorithm and engineering members.

Results:

  • Released Xinyuan-VL-2B, which ranked first on OpenCompass in the under-4B parameter category at the time.
  • Released Xinyuan-LLM-14B-0428, a domain foundation model for mental health and education.
  • Led external technical evangelism, including a keynote at the Alibaba Cloud ModelScope developer event and a joint interview with Alibaba Cloud Tongyi Lab and ModelScope.

Selected Projects

Health LLM Knowledge Platform

Problem: Mental health and medical consultation require high levels of safety, professionalism, and empathy. Deploying general-purpose models directly carries significant hallucination and safety risks.

Solution: Built a health-focused LLM matrix covering foundation models, knowledge enhancement, and application deployment. Processed millions of medical records and hundreds of thousands of psychological dialogues, totaling 4M+ samples. Applied RAG and a three-dimensional memory system to reduce hallucinations.

Results: Incubated Sunsimiao medical LLM (400+ Stars) and MindChat mental health LLM (700+ Stars). Connected model capabilities to WeChat via OXiaoPeng, reaching 2,000+ direct users and 20,000+ indirect users.

ColugoMum: Smart Retail Settlement Platform

Problem: In unmanned retail, products are numerous, visually similar, and frequently updated. Traditional object-detection approaches require retraining for every new product, leading to high maintenance costs.

Solution: Developed an image-retrieval-based product recognition algorithm that eliminates the need for retraining when new products are added.

Results: Won the National First Prize at the 2022 China Robot and Artificial Intelligence Competition, exhibited at Baidu Wave Summit 2021+, and accumulated 100,000+ views and 200+ GitHub Stars.

OXiaoPeng

Problem: Early in the LLM wave, multiple mainstream models were fragmented and hard for ordinary users and small communities to access.

Solution: Built a multi-model aggregation application integrating Baidu ERNIE, PanGu, Yuan1.0, ChatYuan, and ChatGPT, accessible through WeChat, Feishu, and QQ.

Results: Served as a unified access layer for community evaluation and feedback, reaching 2,000+ direct users and 20,000+ indirect users. The project was shortlisted for the 2023 MiraclePlus Spring Camp (S23) interview.

Publications & IP

Awards & Honors

  • First Prize, 2025 Synthetic Data Competition · Lingxi AI for Mental Health
  • Bronze Award, 2024 China International College Students’ Innovation Competition
  • Bronze Award, 2023 9th China International “Internet+” Innovation Competition (Lingxin Intelligence)
  • Silver Award, 2022 8th China International “Internet+” Innovation Competition (Xiaosheng Technology)
  • National First Prize, 2022 24th China Robot and Artificial Intelligence Competition

Selected Honors

  • Baidu PaddlePaddle Developer Expert (PPDE)
  • OpenAtom Foundation Active Open Source Contributor
  • Baidu AIStudio 2022 Top 10 Influential Figures
  • OpenI Community Core Early Experience Officer
  • Datawhale Member
  • Qwen Ambassador

Talks & Media

Invited Talks

  • 2025.09: Speaker, Apsara Conference “Doing AI, the Gen-Z Way”
  • 2025.05: Alibaba Cloud ModelScope Developer Co-Creation Event, “Open Source Technology Driving Pan-Psychological Services and AI Inclusion”
  • 2023.06: PaddlePaddle LLM Application Development Course, “Building an Intelligent Document Query Assistant”
  • 2023.05: Datawhale AIGC Learning Program, “Building a Local Knowledge Base QA Application with LangChain and ChatGLM-6B”
  • 2022.04: PaddlePaddle Industry Practice Library, “Product Recognition Industry Application”

Media Coverage

  • Alibaba Cloud Tongyi Qwen: “Writing Code, Writing Emotions”
  • Founder Park: Interview on Qwen 3 and Cylingo Group
  • Alibaba Cloud Tongyi Qwen: “Tongyi Qwen + Mental Health = ?”
  • ScienceNet: “They Developed a Mental Health LLM with Tongyi Qwen”
  • Synced SOTA Models: Multiple new releases of MindChat and related projects

Community & Activities

PaddlePaddle Navigator Group | East China Regional Lead
2021.09 - 2022.07

  • Oversaw operations across 7 provinces in East China, expanding coverage to 50+ universities including Zhejiang University, Southeast University, ShanghaiTech, Nanjing University of Aeronautics and Astronautics, Soochow University, ECUST, and Hefei University of Technology.

ECUST PaddlePaddle Navigator Group | Lead
2021.04 - 2022.01

  • Built the university-level navigator group from scratch, organizing lectures, competitions, and hands-on projects, reaching 100+ participants and incubating 10+ high-quality projects.
  • Drove the implementation of the Baidu-ECUST industry-education collaboration course.

Education

East China University of Science and Technology | School of Information Science and Engineering | Automation
2019.09 - 2023.06

Contact

I am currently continuing to advance the Xinyuan model series while exploring multi-agent collaboration, embodied AI with affective interaction, and sustainable business models for domain-specific LLMs.

I am open to collaboration in:

  • Domain-specific LLMs for mental health, healthcare, education, and community services
  • Affective intelligence and multimodal interaction products
  • Open source community building and technical evangelism

Reach me at yx20001210@163.com or on GitHub.

← 中文