Hi 👋, I'm Youyao Gao.

         

Smart legislation

• Data Acquisition and Preprocessing: Responsible for collecting corpus data using keyword engineering techniques and performing data cleaning and noise reduction using SQL and Pandas. Utilized intelligent operators to create and optimize data processing workflows, ensuring high quality and availability of data. After preprocessing, the data was randomly divided into training and testing sets, providing a foundation for model training and evaluation.
• Model Selection and Optimization: Selected the BGE-M3 Embedding model, specifically optimized for text and content retrieval. Implemented cross-validation and used grid search and random search methods for hyperparameter tuning, effectively enhancing the model's retrieval performance and identifying the optimal model configuration.
• Data Source Configuration and Workflow Implementation: Configured internal structured data storage using SQL database connection technologies, ensuring effective integration and application of data. Designed and implemented a workflow for generating and managing feature vectors, supporting the functionality requirements of the intelligent search application. After successful workflow execution, participated in the deployment and launch of the intelligent search application.
• Model Evaluation and Performance Testing: Conducted a comprehensive evaluation of the model using an independent test dataset to verify the model's generalization capabilities. Additionally, utilized the developed intelligent search application API to conduct real data retrieval tests, calculating performance metrics such as Hit Rate and Mean Reciprocal Rank (MRR) to quantitatively assess the actual performance of the model.

Smart Q&A for Medical Devices

• Data Preprocessing and Document Parsing: Conducted comprehensive data cleaning and preprocessing on medical device documents, including text normalization, noise reduction, and document structure analysis. Developed and implemented document parsing pipelines to extract and organize relevant information, laying a solid foundation for downstream retrieval and question-answering tasks.
• Model Evaluation and System Optimization: Performed systematic evaluation of the Retrieval-Augmented Generation (RAG) model, utilizing metrics such as answer accuracy and response relevance. Conducted iterative testing and parameter tuning to enhance the overall performance, reliability, and real-world applicability of the intelligent Q&A system for medical devices.

Key Point Evaluation for Embodied Intelligent Robots

• Key-Point Detection and Model Customization: Customized and fine-tuned the OpenPose model from the CMU Perceptual Computing Lab to enable automatic detection and precise extraction of robot joint coordinates from video frames.
• Action Segmentation and Video Processing: Designed and implemented an action segmentation module to automatically identify the start and end frames of target actions within input videos, extracting complete action segments for downstream analysis.
• Trajectory Extraction and Quantitative Evaluation: Developed a key-point trajectory extraction and alignment framework, converting per-frame joint coordinates into time series and applying Dynamic Time Warping (DTW) to compare action trajectories with reference standards. Generated standardized evaluation scores to quantitatively assess the accuracy and consistency of robot actions.

Supervisor: Prof. Huan Jin

• Path Planning Environment Development: Developed a large-scale collision-free path planning framework for urban hybrid truck-UAV delivery systems, addressing UAV battery limits and traffic delays with the Solomon benchmark dataset.
• Reinforcement Learning Algorithm Design: Formulated a custom Markov Decision Process (MDP) with state variables and real-world constraints, and integrated/adapted Proximal Policy Optimization (PPO) algorithms to optimize UAV-truck path coordination.
• Data Pipeline and Model Enhancement: Built data preprocessing pipelines for high-quality RL training datasets, and fused heuristic rules with RL policies to improve model efficiency and scalability.
• Still in progress

Supervisor: Prof. Fiseha Berhanu Tesema

• Model Development and Implementation: Built a small polyp segmentation model on the OpenMMLab framework, integrating a dual-stream CNN-Transformer architecture with a Laplacian pyramid module and boundary-aware loss; implemented and debugged key sub-modules including CNN stream, Transformer stream, edge separation, attention fusion, and loss function.
• Dataset Preparation: Curated and preprocessed the Kvasir-SEG and ETIS-LaribPolypDB datasets, including normalization, annotation screening, and small-polyp subset partitioning.
• Performance Evaluation: Conducted comparative experiments, validating the model’s performance using Dice, IoU, and inference speed metrics.
• Still in progress