I am interested in how LLMs and VLMs reason, plan, use tools, and maintain context over long-horizon tasks.

• Tool-augmented reasoning
• Persistent context and memory
• Multimodal planning and evaluation
• Agent reliability under incomplete observations

I study whether explicit visual geometry can make VLM reasoning more reliable on real images.

• Object-level spatial representations
• Depth and segmentation assisted reasoning
• Spatial relation and physical-size questions
• Separating perception failures from reasoning failures

I build and optimize perception pipelines for dynamic physical environments.

• Object detection and instance segmentation
• Camera calibration and 6-DoF pose estimation
• PnP, Kalman filtering, and state estimation
• CUDA, OpenCV, and real-time CV optimization

Geometry-aware visual reasoning pipeline for evaluating whether object detection, SAM2 segmentation, Depth Anything V2 relative depth, and object-level geometry can improve VLM spatial reasoning.

The project builds a staged benchmark pipeline:

image -> detection -> segmentation -> depth
      -> geometry -> reasoning -> evaluation

Current features include YOLO detection, label normalization, SAM2 mask generation, depth estimation, object-level geometry extraction, reasoning prompt generation, a geometry-only rule baseline, and failure analysis that separates upstream perception errors from reasoning errors.

Real-time deep learning aim bot for Aimlab Sixshot. The system uses screen capture, target detection, and hardware mouse control at 100+ FPS.

Built a custom MiniUNet with approximately 467K parameters for Gaussian heatmap regression, plus an end-to-end workflow from interactive labeling to real-time inference and adaptive flick control.

PyTorch · OpenCV · CUDA

Python · C++ · C# · Java · JavaScript

Linux · ROS · Docker · Git · MySQL · LaTeX

React · Vue · TypeScript · Unity