Email  /  LinkedIn  /  ORCID

I'm an MPhil student in Computer Science and Engineering at CUHK. My work sits between research and engineering: I turn new methods into tested, reusable code, and I check the data before I trust a result. Lately that means LLM agents and their evaluation, agent memory, model serving, and watermarking for generative models.

Currently taking an intentional GAP, I'm open to internships and full-time roles in LLM and agent application engineering and applied AI, in Hong Kong and mainland China.

A customer-service agent for an online shop, built on LangGraph, Milvus and MCP. It routes each question, cites its evidence, and refuses when the evidence is weak; a calibrated gate makes that call, and human review turns refusals into new knowledge. I set the goals and made the design decisions. AI coding agents wrote the code under my specs, tests and review.

In 30 seconds: watch a recorded session / read the decision records / see what broke and how I fixed it.

pip install markdiffusion works today, and I maintain the conda-forge feedstock.

• MarkDiffusion: An Open-Source Toolkit for Generative Watermarking of Latent Diffusion Models. JMLR, 2026. Co-first author. Paper / arXiv / Code
• Can Watermarked LLMs be Identified by Users via Crafted Prompts? ICLR 2025, Spotlight. Second author. arXiv / Code
• Refining Positive and Toxic Samples for Dual Safety Self-Alignment of LLMs with Minimal Human Interventions (PT-ALIGN). IEEE TIFS, 2026. Third author. Paper / arXiv
• RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents. Findings of ACL, 2026. Third author. arXiv / Code
• Omni-SafetyBench: A Benchmark for Safety Evaluation of Audio-Visual Large Language Models. ACM Multimedia 2026, Oral. Co-author. arXiv

MemTensor, algorithm engineering intern (Dec 2025 to Apr 2026). Offline research on storing user preferences in model weights instead of retrieving history.

• Built a three-stage topic, question and answer synthesis pipeline that averaged 160,000+ QA pairs per user, with ROUGE-L filtering, LLM quality checks, concurrent API calls, JSON validation, retries and checkpoint recovery.
• Trained per-user Qwen3-0.6B LoRA adapters and combined them with Qwen3-8B through contrastive decoding. PersonaMem accuracy rose 28 points over unadapted Qwen3-8B, with no user history at inference and a strict train/test split.
• Measured where it broke (unstable gains at larger sizes, loss of general capability), which led the team not to pursue the approach.
• Wrote Chapter 2 of 《记忆工程·智能体记忆系统构建与实践》 and maintain bilingual resources in Awesome-AI-Memory.

LiveServe. Adapted Ming-Flash-Omni to vLLM-Omni as a Thinker, Talker and Code2Wav pipeline, and set the team's TTFP and audio RTF latency baselines.

RecMem. Ran the MemoryOS baseline on LoCoMo and LongMemEval-S: dataset adapters, LLM-as-a-Judge scoring, and token accounting across memory construction and answering.

Course, competition and social-practice projects from BUPT, 2023 to 2024. Each one took me one layer further down or across the stack.

Also: a course-management system for the object-oriented design course (Spring Boot, JPA, H2).

• Agents and LLM apps: LangGraph, LangChain, MCP, Milvus (dense and BM25), rerankers, Langfuse, FastAPI, Playwright
• Models and training: Python, PyTorch, Transformers, Diffusers, vLLM and vLLM-Omni, DeepSpeed, LoRA
• Evaluation and quality: pytest, GitHub Actions, LLM-as-a-Judge, LoCoMo, LongMemEval
• Data and retrieval: Elasticsearch, FAISS, ChromaDB, bge-m3 and rerankers, Ray, FastAPI
• Systems and media: C, x86 assembly, C++, Qt, FFmpeg, OpenCV
• Web: TypeScript, React, Next.js, Three.js, Node.js, Spring Boot, MySQL, MongoDB, RabbitMQ