CV

AI Algorithm Intern

DexForce

Contributed to computer vision model optimization, data processing, and educational robotics deployment in server environments.

• Reconstructed the data augmentation pipeline to mitigate overfitting and adapt models to diverse environments, implementing class-aware masking strategies to preserve image-label alignment.
• Enhanced YOLOv8 architecture for heavily occluded scenes by integrating occ_mask/back_mask auxiliary training and optimizing detection head structures through ablation studies and model pruning, reducing model parameters by 10% while improving multi-occlusion detection performance.
• Built a Mask R-CNN/Keypoint R-CNN-based 6D pose estimation pipeline for educational robotic arms, combining PnP algorithms to solve monocular camera pose estimation, improving recognition accuracy of easily confused objects from 30% to 60%.
• Conducted segmentation model ablation studies to remove redundant convolutional layers and simplify mask generation logic, comparing AP/Recall metrics to identify optimal lightweight solutions.

FinGent: AI-Driven Financial Advisory & Quantitative Backtesting System

• Designed and implemented a multi-agent collaborative financial analysis platform integrating real-time data ingestion, RAG augmentation, quantitative backtesting, and interactive visualization, with agent state persistence and Docker containerization.
• Architected a hierarchical agent workflow via LangGraph spanning intent recognition, data acquisition, risk assessment, expert analysis, and aggregation voting, reducing single-agent context length and mitigating hallucinations through asynchronous concurrent execution.
• Engineered an intelligent data engine using akshare/Tiingo APIs and SEC EDGAR scraping, standardizing interfaces via LangChain Tools for autonomous A-share and U.S. stock queries.
• Implemented PostgreSQL-based checkpointing for agent execution states, supporting crash recovery and time-travel debugging.
• Optimized domain-specific RAG with query rewriting and LLM-powered financial semantic expansion.
• Built a real-time investment research dashboard with Vue3, TypeScript, and ECharts for K-line visualization and agent decision-chain tracking.
• Containerized the complete service with Docker using conda environment export and miniconda base image builds, with data volume mounts for state persistence.

Assistive Vision System for Visually Impaired Individuals

• Architected a multimodal perception system integrating object detection, semantic segmentation, and voice interaction for complex road condition awareness.
• Embedded a SimAM parameter-free attention mechanism into the YOLOv5s backbone to enhance feature extraction while maintaining lightweight model size, achieving 93% accuracy and 92% mAP@0.5 for traffic light and pedestrian detection at over 30 FPS.
• Developed U-Net-based drivable area segmentation with scenario-specific data augmentation, reaching 97% pixel-level accuracy for sidewalk, crosswalk, and obstacle boundary detection.
• Built a high-contrast accessible UI with PyQt5 integrating real-time camera streams and TTS voice alerts, establishing a closed-loop pipeline from visual perception to risk warning with sub-200ms latency.

Second Prize, Blue Bridge Cup Programming Contest (C/C++), Guangdong Division

National Excellence Award, National IT & Software Talent Competition, Digital Technology Innovation Track

Provincial-Level, University Student Innovation and Entrepreneurship Program

University-Level, Double-Hundred Project Team Leader