About
About 371 wordsAbout 1 min
AI engineer, currently pursuing a research Master's (MRes) in Pattern Recognition and Intelligent Systems.
My past work centred on R&D and production delivery in computer vision, OCR, document intelligence, and LLM applications — building systems for document structuring, OCR + LLM, RAG / GraphRAG, Agent Workflow, and model inference deployment.
My research interest now focuses on AI4Math, LLM Reasoning, and Geometry Problem Solving — how models can combine visual information, structured representations, and symbolic tools to reason more reliably over many steps.
I write here about papers I read, things I learn, and engineering practice.
At a glance
| Item | Content |
|---|---|
| Current | MRes Student · AI4Math / LLM Reasoning |
| Previously | 3 years in AI algorithm R&D and technical management |
| Focus | AI4Math · Geometry Reasoning · Document AI · Agentic AI |
| Engineering | CV · OCR · LLM · RAG · GraphRAG · Agent |
Research Interests
Currently focused on:
- AI4Math / Mathematical Reasoning
- Geometry Problem Solving
- Multimodal & Neuro-Symbolic Reasoning
- LLM Reasoning / Test-Time Scaling
- Agentic AI & Efficient Inference
Rather than whether the final answer is correct, I'm more interested in:
Whether the model genuinely understands and makes use of intermediate structure, visual information, and external tools.
Engineering Experience
The past few years centred on Document AI and LLM applications, including:
- Document image preprocessing, layout analysis, and OCR
- OCR + LLM understanding and evaluation
- RAG / GraphRAG knowledge-enhanced Q&A
- LLM SFT / LoRA and inference deployment
- Agent Workflow
- ONNX / TensorRT / vLLM inference optimisation
Representative work includes intelligent document recognition and evaluation, LLM-based assessment, math knowledge graph Q&A, and the personally developed Study Planner AI Agent.
Writing
This site mainly records:
Paper readings · Research notes · AI engineering practice
Across AI4Math, LLM Reasoning, Agents, RAG, computer vision, and inference optimisation.
Rather than just recording how something is implemented, I'd rather answer:
Why is it designed this way? When does it fail? Is there a better approach?
Open Source & Community
- Nature Skills Contributor — contributed the Nature Paper Card Skill
- Take part in GitHub open-source Issue / PR collaboration
- Member of CSIG (China Society of Image and Graphics)
Contact
Happy to talk about:
AI4Math · Geometry Problem Solving · LLM Reasoning · Document AI · Agent Systems
Copyright
Copyright Ownership:Kstheme
License under:Attribution 4.0 International (CC-BY-4.0)