Evaluating single-image layer decomposition methods to build the foundation for a tri-view layered World Action Model in robot manipulation.

A universal back-end compiler parameter optimization system.

A CUPTI-based GPU profiler for LLM inference on SGLang and vLLM, exposing a Python API for kernel-level bottleneck analysis.

An end-to-end image classification pipeline on 20k+ underwater species images, comparing ViT and EfficientNet-B0.