Participating in R&D for Agent RL post-training systems. Designing and implementing low-overhead calibration mechanisms for asynchronous streaming RL training to address data consistency challenges and ensure stability in large-scale RL training.

Developed Pipeline Parallelism (PP) modules for Ascend LLM inference systems. Researched optimization techniques including Parallelism Mechanisms, PD Disaggregation, Speculative Decoding, Sparse/Linear Attention, and Quantization. Addressed resource scheduling and memory bottlenecks in super-node environments.

Researched secure microservice architectures. Designed a secure and efficient microservice runtime based on Trusted Execution Environments (TEE) and Rust.

Efficient PD Disaggregation Inference System Optimized by Pipeline Parallelism (PP)

Huawei Research: Integrated Pipeline Parallelism (PP) & PD Disaggregation to optimize long-sequence and multi-node scenarios. First complete solution combining PP, PD disaggregation, and Multi-Token Prediction. Improved throughput by 9.47% and kept pipeline bubbles ≤5%.

Secure and Efficient Distributed OLAP System via Fully Homomorphic Encryption (FHE)

HKU Research: First FHE-based distributed OLAP system. Achieved secure and efficient data processing. Reduced end-to-end latency by 44.1% compared to SOTA single-node systems in 4-node settings.

High-Throughput OLAP System Accelerated by GPU/NPU

HKU-Huawei Joint Research: Designed high-performance OLAP system using GPU/NPU for offline scenarios. Achieved 50% throughput improvement compared to NVIDIA RAPIDS for Spark.

M.Phil. in Computer Science

B.Eng. in Computer Science and Technology