cheng si yuan

LLM architecture design & inference systems: Decoder-Only evolution (MLA, attention mechanisms), KV-Cache optimization, Prefill-Decode disaggregation (vLLM, NVIDIA Dynamo, DynoSim simulation). Robotics: visual servoing (IBVS, hand-eye calibration, Kalman filter + PID), humanoid RL locomotion (Isaac Sim, PPO, Unitree G1), ROS2. 3D reconstruction: 3D Gaussian Splatting, panoramic SfM, TSDF/Poisson surface reconstruction. Edge AI deployment: model quantization (PTQ/QAT INT8), TensorRT, Jetson/STM32 embedded systems. Multimodal: CLIP retrieval, VLM Agent control, Stable Diffusion.