Applied AI · Multimodal Learning · Intelligent Systems · Agent
I am an MSc Computer Science student at The University of Hong Kong, with research interests in applied machine learning, multimodal representation learning, and AI for real-world intelligent systems.
I am currently a Research Intern at the Tsinghua Intelligent Connected Vehicle Group (THICV). I also contribute to AI research and engineering at Poxiaoyunyi Technology, a BUPT-affiliated technology team.
Previously, I worked as a Data and AI Intern at Lenovo, supporting data analysis, AI-model prototyping, and product-oriented technical research.
| Project | Engineering focus | Stack |
|---|---|---|
| UK Visitor Visa Agent | Local-first document review agent that checks dates, travel plans, finances and sponsorship evidence, then produces risk flags, checklists and draft explanations. Ships as a web app, Windows desktop release and reusable Codex skill. | TypeScript, Next.js, OpenAI API, Electron |
| AI Teaching Agents | Team-built with classmates: a multi-agent teaching system for personalized content, intelligent practice, learning analytics, document processing and session management, with support for OpenAI, DeepSeek, Qwen and custom model providers. | Python, Streamlit, Agno, LLM APIs |
| Cantonese Speech Emotion Recognition | Graduation thesis project for Cantonese speech-emotion recognition, combining multimodal features with classical and neural baselines, reproducible training pipelines and experiment reports. | Python, PyTorch, librosa, scikit-learn |
| Urban Scene Semantic Segmentation | U-Net and DeepLabV3+ pipelines for dense urban-scene understanding. The retained U-Net experiment records 0.6786 mIoU and 89.42% pixel accuracy on its validation split. | Python, TensorFlow, PyTorch, OpenCV |
| Commerce Operations Platform | Full-stack operational system spanning product data, marketplace orders, procurement, inventory, warehouse fulfilment, settlement and reconciliation. Published as a sanitized engineering snapshot with no business data or credentials. | Next.js, TypeScript, React, Prisma, SQLite |
| CIFAR Robust Training Pipeline | Reproducible image-classification workflow covering data materialization, augmentation, robust training, checkpoints and metric-driven evaluation. | Python, PyTorch, torchvision, scikit-learn |
| MSc Buddy | Cross-platform collaboration workspace for team formation, project planning, task ownership and deadline tracking. | Flutter, Dart, Provider |
I contribute fixes and tested improvements to external repositories. The status below is stated explicitly so merged work and open work are not conflated.
| Project | PR and status | Contribution |
|---|---|---|
| The PR Agent | #2863 · Merged | Made partial /improve analysis visible: tracks failed and total chunks, reports incomplete suggestion coverage and adds regression tests across complete, partial, fallback, disabled and artifact-output paths. |
| The PR Agent | #2867 · Open | Corrects fallback chunk alignment after a conversion failure and adds regression coverage for changed chunk boundaries. |
| Prowler | #12650 · Open | Extends Cyber Essentials 3.3 compliance mappings to AWS while preserving Azure mappings, with provider-discovery and compliance-output tests. |
- Multimodal perception, autonomous systems and sensor fusion
- Semantic segmentation, image classification and signal-based learning
- Reproducible experimentation, evaluation and evidence-backed reporting
- Full-stack operational systems, data workflows and deployment engineering
ML and vision: Python, PyTorch, TensorFlow, OpenCV, scikit-learn
Systems and applications: TypeScript, Next.js, React, Flutter, Java, C/C++
Data and operations: Prisma, SQL, Linux, Nginx, Git
- KidsBank - team-built JavaFX financial-literacy application with structured controllers and tests
- ParkShare - responsive parking-sharing product concept with driver, owner and wallet flows
- Obstacle-Avoidance Robot - Arduino control system using ultrasonic sensing, motor actuation and an I²C display
Public repositories are curated to keep claims traceable, source readable and private datasets, credentials and operational records out of version control.
应用机器学习 · 多模态感知 · AI Agent
我是香港大学计算机科学硕士研究生，研究方向为应用型机器学习、多模态表征学习，以及面向真实场景的智能系统人工智能技术。
目前任职于清华大学智能网联汽车研究组（THICV）担任研究实习生，从事端到端自动驾驶相关工作。同时，我也在北邮团队破晓云逸科技负责人工智能科研与工程落地工作。
此前，我曾在联想担任数据与人工智能实习生，负责数据分析、AI 模型原型开发以及面向产品的技术调研。
| 项目 | 工程重点 | 技术栈 |
|---|---|---|
| 英国访客签证材料助手 | 本地优先的材料审查 Agent，核对日期、行程、资金与担保材料，输出风险提示、材料清单和说明草稿；同时提供 Web 应用、Windows 桌面版本和可复用 Codex skill。 | TypeScript, Next.js, OpenAI API, Electron |
| AI Teaching Agents | 与同学共同开发的多 Agent 教学系统，支持个性化内容生成、智能练习、学习分析、文档处理和会话管理，并兼容 OpenAI、DeepSeek、通义千问及自定义模型服务。 | Python, Streamlit, Agno, LLM APIs |
| 粤语语音情感识别系统 | 毕业设计项目：面向粤语语音情感识别，将多模态特征与经典机器学习及神经网络基线结合，提供可复现训练流程和实验报告。 | Python, PyTorch, librosa, scikit-learn |
| 城市街景语义分割 | 面向城市街景理解的 U-Net 与 DeepLabV3+ 训练流程；保留的 U-Net 实验在验证集上达到 0.6786 mIoU 与 89.42% 像素准确率。 | Python, TensorFlow, PyTorch, OpenCV |
| 电商运营管理平台 | 覆盖商品、平台订单、采购、库存、仓储履约、结算与对账的全栈业务系统；公开版本已移除业务数据和凭据。 | Next.js, TypeScript, React, Prisma, SQLite |
| CIFAR 鲁棒训练流程 | 可复现的图像分类流程，涵盖数据生成、增强、鲁棒训练、检查点管理与指标驱动评估。 | Python, PyTorch, torchvision, scikit-learn |
| MSc Buddy | 面向组队、项目规划、任务归属和截止日期管理的跨平台协作空间。 | Flutter, Dart, Provider |
我向外部开源仓库提交经过测试的缺陷修复和工程改进。下表明确区分已合并成果与仍处于开放状态的工作。
| 项目 | PR 与状态 | 贡献内容 |
|---|---|---|
| The PR Agent | #2863 · 已合并 | 修复 /improve 在部分分析分块失败时无法体现覆盖范围的问题：记录失败及总分块数、展示不完整覆盖提示，并为完整、部分、回退、关闭配置和产物输出路径补充回归测试。 |
| The PR Agent | #2867 · 开放 | 修复转换失败后回退分块的成员与边界错位问题，并为分块数量变化补充回归测试。 |
| Prowler | #12650 · 开放 | 在保留原有 Azure 映射的基础上，为 Cyber Essentials 3.3 增加 AWS 合规检查映射，并补充 provider discovery 与合规输出测试。 |
- 多模态感知、自动驾驶系统与传感器融合
- 语义分割、图像分类与信号学习
- 可复现实验、模型评估与证据导向的技术报告
- 全栈业务系统、数据工作流与部署工程
机器学习与视觉： Python、PyTorch、TensorFlow、OpenCV、scikit-learn
系统与应用开发： TypeScript、Next.js、React、Flutter、Java、C/C++
数据与运维： Prisma、SQL、Linux、Nginx、Git
- KidsBank - 团队开发的 JavaFX 金融素养应用，包含结构化控制器与测试
- ParkShare - 响应式停车位共享产品原型，包含车主、驾驶员与钱包流程
- 避障机器人 - 基于超声波感知、电机控制与 I²C 显示屏的 Arduino 控制系统
公开仓库经过整理，确保项目描述可追溯、源码可读，同时不会将私有数据、凭据或运营记录提交到版本控制中。