Developed a multi-agent system using Google ADK and Gemini 2.5 Flash Lite to autonomously convert natural language hardware specifications into verified SimPy simulation code
Achieved an 87x reduction in development time (from hours to ~41 seconds on average) by automating the translation and verification workflow with high precision
Designed a specialized three-agent architecture (Analyst, Architect, Verifier) that decomposes complex engineering tasks into structured analysis, code generation, and rigorous validation
Implemented a "Self-Healing Loop" where the Verifier agent autonomously detects simulation errors and provides feedback to the Architect for iterative code correction without human intervention
kDEN-NSGA-II Scheduler for Flexible Job-Shop Scheduling, 06.2025
Developed an AI-driven health consultation platform using LLM and RAG, supporting text and voice interactions with PubMed paper retrieval, along with a CLI-based database management system
Designed a scalable 4-layer system architecture (Process Centric, Business Logic, Adapter Services & Data Services) for efficiency and security
Utilized FastAPI for the backend, SQLite for data storage, Pinecone for vector search, OAuth2.0 with JWT for authentication, and PubMed API integration
Built a responsive frontend with HTML, CSS, and JavaScript, ensuring seamless user interaction via text and voice inputs
Heuristic NSGA-II for Flexible Job Shop Scheduling Problems, 02.2025
Developed an advanced NSGA-II algorithm with heuristic population initialization for solving the Flexible Job Shop Scheduling Problem (FJSP), focusing on multi-objective optimization to balance makespan and workload.
Designed a modular project structure, including data preprocessing, algorithm implementation, batch processing, evaluation metrics, and visualization
Conducted experimental validation on benchmark datasets, demonstrating performance improvements in scheduling efficiency.
High Performance Computing for Grey Wolf Optimizer (GWO) Optimization, 01.2025
Designed and implemented the HGT-GWO algorithm, incorporating global historical best positions and individual trend guidance, significantly improving convergence speed and outperforming traditional GWO on three benchmark functions
Proposed a novel master-worker island parallelization scheme, enabling independent subpopulation operations and reducing communication overhead through controlled synchronization intervals, thereby enhancing parallel efficiency
Conducted experimental validation of the HGT-GWO algorithm using Python, demonstrating superior performance over GWO on 15 test functions
Developed a fully parallelized implementation utilizing C, MPI, and OpenMP, tailored for UNITN's HPC cluster to optimize computational resources
Augmented Reality-Driven Robotic Arm Control for Industrial Automation (SUPSI Project Course), 09.2024
Conducted a comprehensive literature review on XR technologies in Industry 5.0, highlighting human-centric design, worker safety, and data privacy issues [PDF]
Developed and tested a system that combines YOLOv8 and FastSAM models to achieve accurate image segmentation and fingertip coordinate mapping
Designed an AR-based interface using Microsoft HoloLens 2 for real-time gesture recognition, allowing intuitive robotic arm control
Achieved highly efficient performance in industrial environments, mitigating challenges such as hand occlusion using time-sharing processing and ensuring flexible task handling
Demonstrated scalability through multi-mode operation, enabling both gesture-based and interface-based controls for part picking
Research on Semantic Segmentation Method of High-Resolution Remote Sensing Images Based on Non-Local Attention Mechanism with Deep Learning, 05.2023
Conducted comprehensive research on efficient and accurate image segmentation algorithms for complex remote sensing images
Developed an encoder-decoder model with residual-weighted attention to enhance feature extraction and mitigate performance degradation
Conducted experiments on ISPRS Potsdam and Vaihingen datasets, achieving F1-scores of 90.23% and 87.37%, respectively, outperforming models without attention mechanisms
Demonstrated proficiency in convolutional neural networks, Transformer models, and attention mechanisms for semantic segmentation tasks
Video anomaly detection based on unsupervised learning, 12.2022
Reproduced the paper HF^2-VAD(ICCV 2021, SOTA), tested on CUHK Avenue dataset, the performance is slightly better than that of the paper (0.2%)
Music Generation Toolkit, 12.2022
Invited by NLP teacher to give a lecture to the next cohort of students(her new students)
A collection of excellent music generation models in recent years. The music data format includes compound word and REMI. The model is mainly transformer, including transformer XL, Vanilla Transformer, etc. It can freely combine models to generate music.
All are developed on the Jetson Nano, including data collection, model training and testing
Support real-time data reading (captured by camera) and automatically give game results
Automatic compensation of agricultural insurance based on blockchain technology and remote sensing(College Students' innovation and entrepreneurship training program), 06.2022
Judge the disaster situation of crops through remote sensing images (image segmentation)
Use blockchain technology to encrypt the insurance compensation process
Music Generation Based on LSTM, 11.2021
Used music21 to complete the conversion of midi format, and use LSTM to build a music generation model