Machine Learning Researcher · Agent Builder · Open-source Contributor
- 💡 I’m a serial entrepreneur exploring AI × Hardware. My current venture focuses on AI-driven sleep intervention and was selected for incubation within Prof. Zexiang Li’s entrepreneurship ecosystem behind DJI and XbotPark, with RMB 500K in initial funding.
- 🎓 M.S. student in Software Engineering at Southeast University (QS World University Rankings 2026: #392), previously B.Eng. in Computer Science at Hohai University (QS World University Rankings 2026: #1001–1200).
- 🧠 My work focuses on Agent Memory, Agent Systems, and Learning under Imperfect Data.
- 📄 Research interests include long-tail learning, partial-label learning, noisy supervision, and robust machine learning.
- 🛠️ Contributor to open-source AI projects including QwenPaw, ms-swift, LanceDB, Docling, Feast, and more.
- 🚀 I enjoy building across research, engineering, and real-world deployment.
- 🤝 Open to collaboration on Agent, LLM, recommendation, and ML systems.
- 📫 Reach me at: yaodong.shen@seu.edu.cn
Underlying models will keep evolving — from today’s LLMs to future World Models and beyond — but the core problems around Memory, Context, Tool Use, and Runtime will remain fundamental to intelligent agents.
Model paradigms will continue to change, but a solid foundation in Machine Learning determines how quickly one can understand, adapt to, and build with new technologies.
| Layer | What I work on | Connected open-source projects |
|---|---|---|
| Agent Systems | Agent building, multi-agent collaboration, real-time multimodal interaction | QwenPaw · LiveKit Agents · agentUniverse |
| Memory & Retrieval | RAG orchestration, local-first memory, document understanding and retrieval | Haystack · AutoRAG · Basic Memory · LanceDB · Docling |
| Learning & Evaluation | Partial-label learning, long-tail robustness, model training and evaluation | ms-swift · LM Evaluation Harness |
| ML Systems Foundation | Reusable data and feature infrastructure | Feast |
| Work | Focus | Current status |
|---|---|---|
| AdapMatch | Adaptive bias decoupling for semi-supervised partial-label learning under unknown class distributions | Under review · NeurIPS 2026 · Student first author |
| Three Stones, One Bird | Generalization-error-bound-guided noisy partial-label learning | Under review · TPAMI 2026 · Student first author |
Every entry below links to a merged pull request. I keep the scope broad deliberately: the projects form the systems map above, rather than a collection of isolated patches.
| Project | Contribution area | Merged PR |
|---|---|---|
| Docling | Document parsing and structured conversion | #3954 |
| QwenPaw | Personal AI agent and extensible skills | #6502 |
| Haystack | LLM-agent and RAG orchestration | #12218 |
| ms-swift | LLM and multimodal training | #9832 |
| LM Evaluation Harness | Language-model evaluation | #3959 |
| LiveKit Agents | Real-time voice and multimodal agents | #6501 |
| LanceDB | Multimodal retrieval database | #3699 |
| Feast | Feature-store infrastructure | #6697 |
| AutoRAG | Self-evolving retrieval agent | #1410 |
| Basic Memory | Local-first Markdown memory and knowledge graph | #1348 |
| agentUniverse | Expert-collaboration multi-agent framework | #832 |
Profile design: a systems-layer view inspired by open-source AI ecosystems. Facts and links are maintained in this repository.