💼 Experience
Tencent · WeChat Group (Project-Up, Talented Program)
Research Intern · LLM agentic post-training
- Agentic dialogue system (in production): a conversational product-selection agent for creators on a large-scale in-house short-video commerce platform, evolved from a fixed LLM workflow to a ReAct agent with failure recovery, then to a skill-based architecture.
- Agentic post-training & compression (in production): sketch-of-thought SFT distilling Hunyuan3 into Qwen3.5-9B; per-GPU capacity rose from 3 to 64 concurrent users (21×), enabling deployment at up to 2,000 users.
- Silent tool-error benchmark: RITE, evaluating how agents identify, re-fetch and recover from plausible but wrong tool outputs.
NVIDIA · Solution Architecture, Hong Kong
Solutions Architect Intern (Research, Part-time) · Machine learning infrastructure
- CUDA-Q ML (ISC'26 poster): makes quantum circuits differentiable nodes in PyTorch autograd, with automatic gradient selection and distributed data-parallel training.
- Tile-granularity fast inference engine: suspendable decoder tiles scheduled from one queue cut urgent QEC decoding turnaround by 3.6–16.6× at only 0.2–1.1% throughput cost.
- LLM training & inference parallelism: Megatron-LM and vLLM, plus technical Q&A for NVIDIA's partner universities working on AI4Science.
HKUST (Guangzhou) · Information Hub
Funded Research Intern · Data science and analysis
- ParaGAN, a parameter-controlled adversarial generative model for cervical cell image augmentation (Med Data Lab).
👨🏭 Projects
VR English Program
Audio Module Enhancement with Audio Language Model Adaptation for Secondary Students English Education in VR and Physical World
Center for Language Education @ HKUST & VisLab @ HKUST. Part of the “AI-assisted Virtual Reality English Speaking Program for Secondary Students”.
- 30 million HKD from the Quality Education Fund (HK), 9 million for us and 1.4 million HKD for this part
- Down-stream task adaptation and service deployment: built a small SFT dataset from speech scores and distilled it into Qwen2-Audio, Qwen2.5-Omni and traditional models (e.g., Wav2Vec), reaching Spearman and Pearson correlations above 0.8 on secondary-student data. Developed a stream-based service system to organize distributed requests, supporting up to 300 simultaneous users on a single 80G A800.
AoE Digital Twin
Guangdong-Hong Kong-Macao Great Bay Area Regional Earth System Digital Twin Project
- 60 million HKD from RGC@HK, 1.4 million for us and 0.4 million HKD for this part
- Responsible for “Neural Network and Distillation Based Ocean Data Compression for Visualization Acceleration”. Developed a machine learning-based compression framework for real-time, interactive visualization of large-scale scientific data, compressing daily data files from over 6 GB to a 6.75 MB model and cutting processing and rendering time to 0.75 seconds on an Nvidia 3070.
- The pipeline is inspired by Neural Radiance Fields (NeRF): tiny neural networks compress each time step spatially, and knowledge distillation then aggregates the temporal models into a single compact model for edge deployment.