I am an MSc student at the University of Hong Kong, working on AI. My interests include multimodal models, LLM agents, long-video understanding and generation, and efficient video inference. I enjoy turning complex ideas into clear, practical learning resources.
HONG KONG · SYDNEY / CURIOSITY, MADE PRACTICAL
Practical workspaces with inspectable evidence and clear limits.
把客服大模型做成能上线的系统：回答逐句对照知识库和业务数据校验，超额退款暂停等人工审批，模型故障时转人工或引用政策原文，坐席的每一次纠错都会变成知识补丁和回归测试。
FastAPI + React 客服控制台
RAG、订单与退款工具、人工审批
审计日志、链路追踪与数据飞轮
在线演示是控制台的静态构建，数据录制自真实后端的离线运行（虚构数据），可以切换顾客、坐席、管理员浏览。完整版本在本地运行，可接入豆包、混元、DeepSeek 等模型。
两个租户回归门禁证据可追溯自托管 SSDLite MobileNet V2
本地图片与短视频分析
可恢复工作区与中文交接单
无需账户或 API key，素材在浏览器内处理。通用物体检测与场景规则产生待核实线索；专项违规行为识别与真实场地效果仍待验证。
真实小模型证据可追溯Check whether reported profit has cash-flow support. Compare six real companies, inspect AI anomaly cues, test operating and funding scenarios, and export an evidence-linked research brief.
Official SEC annual accounts
Historical-profile anomaly detection
Browser stress scenarios and export
Use the online workspace without installing Python. Real annual filings and Treasury rates have explicit source dates. A separate synthetic lab explains forecasting and information timing.
Six real companiesSource-linked briefsExplore a course, follow its knowledge map, and connect intuition to the mathematics.
Start with your first line of Python. Work through small examples, practise on your own, and build four complete projects. This handbook is in Chinese.
19 learning units · 105 runnable examples
114 self-study exercises · 4 projects
Search chapters, follow visual demonstrations, and save your progress in your browser.
Python 3.10+Standard libraryOffline readingFollow values, objects, contracts and collections from source code to a working application.
Reason about bytes, ownership and coordination from a C program to a parallel system.
Choose an algorithm, trace its steps, prove its rule, and account for its cost.
From finite memory to the limits of algorithms, with logic as a language for precise reasoning.
Turn search spaces, observations and examples into carefully justified decisions.
Bring the core formulas, models, and decision rules together in one printable collection.
23 additional guided cases connect the existing lessons to practical decisions. Build safer C programs, test algorithm choices, work through proofs, and check AI results. Every case includes a story, complete code, expected output and step-by-step reasoning in English and Chinese.
A familiar situation gives an abstract idea somewhere to begin.
Explain the mechanism, its assumptions, and where the analogy stops.
Read each symbol and connect the equation to its meaning.
Trace concrete steps, then test your understanding with a fresh question.
Learn one idea at a time. Start with a story. Follow the maths. Run Python and check its output. Each guide moves from the basics to common applications.
Begin with eight short lessons on Python, vectors, probability and gradients. No prior AI knowledge is needed.
From a trustworthy question to a defensible model
Follow the story ↗From tensor shapes to adaptable representations
Follow the story ↗Understand images, geometry, recognition and the evidence behind a prediction.
Follow the story ↗Connect pictures, words, sound, and time with evidence
Follow the story ↗Independent bilingual companions to official tutorials and original research from Google, PyTorch, scikit-learn, Stanford, OpenCV, Hugging Face and the research community. Every topic links to its sources.
Long videos require answers that can be checked against visual evidence. This work introduces EV²-Bench and DynamicSelect to study evidence-grounded understanding and efficient video processing.
Read the publication ↗My contributions included benchmark construction, temporal and visual evidence annotation, model evaluation, and evidence scoring using temporal IoU, text similarity, and bipartite matching.
Long videoEvaluation
I use stories, small examples, and careful definitions to make technical ideas easier to revisit.
EDUCATION
MSc
Working on AI. Enrolled September 2026.
Studied Finance and Computer Science
Dalyell Scholar.
Exchange study
Short-term exchange experience.
Mandarin Chinese and English (IELTS 7.5)
EXPERIENCE
EV²-Bench & model evaluation
Contributed to dataset construction and reproducible evaluation with lmms-eval, structured model outputs, and evidence precision, recall, and F1.
Shenzhen Beidou Applied Technology Research Institute
Worked on video data preparation, annotation consistency checks, and model fine-tuning, distillation, quantization, and deployment validation for safety monitoring.