Major Projects
Done
QA-FineTune Project
LLM chatbot project that introduced MCP-based model calls and a vLLM serving pipeline, then used CoT and DeepEval Synthesizer golden data plus SFT to improve benchmark performance by up to 22% and response speed by 66%.
View Details →
- Provided a chatbot idea to VS Community.
- Introduced a Model-Context-Protocol (MCP)-based model invocation interface and a vLLM serving pipeline to separate model-specific runtime dependencies and reduce response-latency bottlenecks through asynchronous parallel inference.
- Built a Chain-of-Thought (CoT) and DeepEval Synthesizer-based golden dataset for complex query scenarios, then applied Supervised Fine Tuning (SFT) to reduce mismatches in question-answer formatting, improving benchmark performance by up to 22% and response speed by 66%.
Done
RoboDine
Vision and robotics project that improved 6-DoF pose-estimation accuracy by 14% with YOLO pose training, implemented Hough Circles and DBSCAN-based foreign-object logic, and collaborated through GitHub, Jira, and Confluence in an Agile workflow.
View Details →
- Trained a YOLO pose model with data preprocessing and augmentation to reduce pre-grasp pose-estimation errors for a robot arm with a monocular 2D camera, improving 6-DoF pose-estimation accuracy by 14%.
- Implemented Hough Circles and DBSCAN-based logic to prepare for scenarios where foreign objects cause issues in the loading area.
- Collaborated through GitHub, Jira, and Confluence while following an Agile development process.