CV
Education, research focus, experience, and selected project work.
Contact Information
| Name | Kai Ding |
| Professional Title | MCS Student | Efficient Vision Transformers and AI Systems |
| 2531046278@qq.com | |
| Phone | +86 13783798601 |
| Location | Sydney, NSW |
| Website | https://dingkai0321.github.io |
Professional Summary
Master’s student in Computer Science at the University of Sydney, advised by Dr. Karlos Ishac, with research focused on efficient Vision Transformer architectures and project experience spanning multimodal RAG, compliance automation, agent infrastructure, and context governance.
Experience
-
2025 - 2026 Beijing, China
AI Engineer Intern
Lenovo (Beijing) Co., Ltd.
Designed and developed a multimodal RAG compliance review system for Lenovo FIFA marketing assets.
- Built a rule-extraction pipeline for Lenovo Brand World pages and FIFA guideline PDFs with PyMuPDF, PaddleOCR, and multimodal models.
- Converted co-branding constraints into atomic rules with source page and coordinate traces, then implemented hybrid retrieval with BGE-M3, Milvus, RRF, and BGE-Reranker.
- Improved Recall@10 from 76.8% to 92.4% and answer faithfulness to 95.1% for rule-grounded retrieval.
- Orchestrated logo detection, spacing measurement, color analysis, retrieval, and compliance judgment with FastAPI and Pydantic, reducing single-image P50 latency from 74.2s to 32.8s.
Education
-
2024 - Present Sydney, Australia
Master of Computer Science
The University of Sydney
Efficient Vision Transformer Architectures
- Research pathway focused on adaptive token compression and efficient Vision Transformer design.
- Advised by Dr. Karlos Ishac.
-
2020 - 2024 Tianjin, China
Bachelor of Engineering
Tianjin University of Technology
Computer Science and Technology
Publications
-
2026 Class-Attention-Guided Adaptive Token Clustering for Efficient Vision Transformers
The University of Sydney
Master’s thesis on adaptive token clustering for Vision Transformers, supervised by Dr. Karlos Ishac, using class attention, adaptive token budgets, and feature-position clustering to compress redundant visual tokens while preserving classification performance.
Skills
Languages
Interests
Projects
-
Otto Personal AI Assistant
A personal AI assistant with CLI, web dashboard, and voice entry points, designed for long-horizon daily work and research tasks.
- Built long-term memory and Knowledge RAG with PostgreSQL, pgvector, full-text retrieval, HNSW indexing, and retrieval gating.
- Added progressive Skill loading and MCP-based tool extension so external tools can be registered, governed, and invoked through a unified runtime.
- Implemented layered context governance with result archiving, observation compression, message trimming, trace collection, and permission-aware tool execution.
-
FIFA Brand Compliance Review System
An industry-facing multimodal compliance system for co-branded Lenovo FIFA marketing materials.
- Combined brand-rule parsing, hybrid retrieval, visual detection, and structured compliance judgment into an end-to-end review workflow.
- Delivered rule-level traceability with linked rule IDs, source pages, quoted standards, and annotated outputs for auditability.