My Portfolio

Welcome to my portfolio!
Below are selected research, projects, and competitions I’ve completed. Click on any title for details.

Experience

Adaptive Model Selection for Real-Time Heart Disease Detection

Designed an AMS framework with an Anytime CNN for real-time 5-class cardiac severity classification; 91.5% accuracy with 0 deadline misses on Raspberry Pi. Published at IEEE RTCSA 2025 (2nd author).

Cross-Modal LLM-Based Robotic Arm Interaction and Control System

Designed an LLM agent with a tool-calling interface to translate natural-language instructions into ROS action sequences; validated full sim-to-real stack on JAKA Zu5 with MoveIt.

Competitions

National Undergraduate Electronic Design Contest

Built a vision-guided pan–tilt auto-tracking system with discrete-time PID and Kalman filtering; <1 cm positioning error and 0.5–0.6 s response time.

National Undergraduate Embedded Chip and System Design Competition

Designed a collaborative quadcopter control system using STM32F4 and NRFx, achieving stable multi-drone motion synchronization.

Mathematical Contest In Modeling

Applied ARIMA, BP neural networks, and K-means clustering to model and forecast trends in board game design.

SUSTech Electronic Design Competition

Built a smart autonomous medicine delivery vehicle with vision-based lane detection and trajectory control on an embedded platform.

Projects

IDSS: Interpretable Diving Action Quality Assessment Platform

Built an interpretable AQA framework for competitive diving using pose-driven auxiliary supervision; improved tIoU@0.75 by +5.68% over the FineDiving baseline and converged 6× faster.

Cloud-Native IoT Network Security Analysis Pipeline

Built a 6-class intrusion detection pipeline on GCP Dataproc/Spark over 20M MQTT records; Random Forest (78.52%) outperformed MLP variants; deployed with Kafka streaming.

Intelligent Scissors: Interactive Image Segmentation Tool

Implemented an interactive Live-Wire image segmentation tool in Java with Dijkstra/A* seed-caching and UX enhancements; 5.2×–11.2× faster than Photoshop with IoU 0.88–0.97.

Reinforcement Learning for Inverted Pendulum Control

Applied Q-Learning, DQN, PPO, and DDPG for inverted pendulum swing-up and stabilization; 100% success rate under ideal conditions and 90% under noise.

eMeritBox

Built a gravity-sensing donation box on Raspberry Pi integrating gyroscope tilt control, a MAX7219 matrix display, and a Flask web interface for remote operation.

Statistical Learning for Data Science

Applied hybrid pipelines (frozen ResNet18 features + classical classifiers and fine-tuned CNN) for fundus lesion classification; 100% test accuracy with a 2.2× faster custom lightweight CNN.

ROS Robot Intelligent Navigation and Control System

Developed a full ROS2 autonomous navigation stack with custom A* global planner and trajectory feedback controller; zero collisions, tied 1st in final course challenge.

Comparative Analysis of Clustering and Dimensionality Reduction Techniques

Implemented and compared K-means variants, PCA, and autoencoders for clustering and dimensionality reduction on real-world datasets.

Classification Using Logistic Regression and MLP

Built and compared binary classification models using Numpy-based logistic regression and MLP; analyzed convergence behavior and generalization performance.