Zitao Liao

Research Experience

Sun, 19 Apr 2026 00:00:00 +0000

This page is now an index. Each research item has been split into an independent post:

Projects

Sat, 18 Apr 2026 00:00:00 +0000

This page is now an index. Each project has been split into an independent post:

【Generative AI】Weibo Personality Analyzer
GitHub: Public repository link not provided
【Operating System】ProcFS Filesystem Based on Rust & Asterinas
GitHub: Tsurumalu/OS-ProcFS-Project
【Deep Learning】Intelligent Sanitation Robot with Deep Learning
GitHub: EpsilonZYJ/NUS-SOC
【Deep Learning】Real-time Facial Emotion Analysis on Jetson Nano
GitHub: Q-Daisy/Ultra-Light-Fast-Generic-Face-Detector-1MB
【MCM】Modeling and Optimization of Voting Mechanisms
GitHub: Public repository link not provided
【Computer Organization】Pipelined RISC-V CPU Design in Verilog
GitHub: Diaosi1317092/Computer-Organization-Project
【Database】openGauss vs. PostgreSQL Performance Benchmark
GitHub: Tsurumalu/DB-Project3
【Java Programming】Java Swing Match-3 Puzzle Game
GitHub: Tsurumalu/3-Match-Java
【Digital Logic】Range Hood FPGA Control System in Verilog
GitHub: Diaosi1317092/Digital-Logic-Project
【Machine Learning】Loan Default Prediction with Machine Learning
GitHub: Public repository link not provided
【Computer Network】P2P Reliable File Transfer with Congestion Control
GitHub: Tsurumalu/Computer-Network-Project-P2P
【Distributed System】SUSTech Merch Store Microservices
GitHub: Tsurumalu/SUSTech-Merch-Store

Competition on LLM Designed EA, GECCO

Sun, 01 Mar 2026 00:00:00 +0000

背景

时间：Mar 2026
场景：GECCO 相关竞赛，围绕 LLM 设计进化算法。

方法

基于 LLM4AD 与 Evolution of Heuristics (EoH) 设计自适应多算子进化算法。
引入 Multi-Armed Bandit (MAB) 模块，用于优化过程中算子自适应选择。

结果

在 GNBG 基准问题中达到 17/24 的最优结果。

技术栈

Python
LLM4AD
Evolutionary Algorithm
Multi-Armed Bandit

反思

后续可补充不同算子调度策略在不同问题族上的泛化对比。
可以增加消融实验，量化 MAB 模块的边际收益。

Modeling and Optimization of Voting Mechanisms

Sun, 01 Feb 2026 00:00:00 +0000

Background

Time: Feb 2026
Context: Competition project for 2026 MCM.

GitHub Repository

Public repository link was not provided for this project.

Highlights

Addressed voting controversies in “Dancing with the Stars” using a Bayesian inverse estimation framework.
Built counterfactual simulations and an XGBoost model under the CPISeq framework.
Proposed a “Four-Strike Rank Fusion” voting rule to improve fairness in high-conflict scenarios.

Tech Stack

Python
Bayesian Modeling
XGBoost
Counterfactual Simulation

P2P Reliable File Transfer with Congestion Control

Fri, 19 Dec 2025 00:00:00 +0000

Background

Time: Dec 2025
Context: Group course project for SUSTech CS305 (Computer Networks).
Repository: Tsurumalu/Computer-Network-Project-P2P

Goal

Build a reliable peer-to-peer (P2P) file transfer application on top of UDP. All transport semantics—handshaking, reliable delivery, retransmission, and congestion control—are implemented at the application layer. The system uses a network simulator for testing under packet loss, varying topologies, and peer crashes.

System Overview

File & chunk model: Files are split into 512 KiB chunks; each chunk is identified by a SHA-1 hash (20 bytes).
Peer model: Each peer holds a fragment (subset) of chunks and can upload to or download from other peers concurrently.
Download flow: On a DOWNLOAD command, a peer discovers missing chunks via WHOHAS / IHAVE, requests them with GET, then reassembles completed chunks into a pickle-serialized fragment file.
Concurrency: Single-threaded event loop using select; multiple chunk downloads from different peers run in parallel, with per-session upload limits enforced via DENIED.

Packet Protocol

Custom UDP packets (max 1400 bytes) with header fields for type, length, and sequence/ACK number:

Type	Code	Role
WHOHAS	0	Query which peers hold requested chunks
IHAVE	1	Reply listing available chunks
GET	2	Request a specific chunk
DATA	3	Transfer chunk payload segments
ACK	4	Acknowledge received segments
DENIED	5	Reject when upload connection limit is reached

Implementation Highlights

Reliable data transfer (RDT): Per-packet sequence numbers and ACKs; dynamic timeout from RTT estimation (α = 0.15, β = 0.3); timeout-based and fast retransmit (3 duplicate ACKs).
Congestion control (Tahoe-style): Slow start and congestion avoidance with cwnd and ssthresh; loss triggers window halving and return to slow start. cwnd history is recorded and plotted with matplotlib for presentation.
Structured codebase: Refactored peer.py with PeerState, DownloadSession, UploadSession, and a Peer class driving a non-blocking I/O loop.
Robustness: Handles concurrent multi-peer downloads, connection limits, peer crashes, and out-of-order segment buffering.
Testing: Validated with course-provided pytest suites (handshaking, basic transfer, concurrency, crash scenarios, and advanced topology tests) via grader.py and the network simulator.

Tech Stack

Python 3.12
UDP (socket) with select for single-threaded concurrency
SHA-1 chunk hashing
matplotlib (congestion window visualization)
pytest & network simulator (grading and integration tests)
Linux / WSL (required runtime environment)

ProcFS Filesystem Based on Rust & Asterinas

Mon, 01 Dec 2025 00:00:00 +0000

Background

Time: Dec 2025
Context: Main practical assignment for the course “Operating System(H)”.

GitHub Repository

https://github.com/Tsurumalu/OS-ProcFS-Project

Highlights

Designed and implemented a ProcFS virtual filesystem for an OS kernel using Rust and the Asterinas framework, with Cargo as the build tool.
Built ProcFS support for system info files, core syscalls such as getdents64 and newfstatat, per-process details including environ, cmdline, fd, and status, and VirtIO device information.

Tech Stack

Rust
Asterinas
Cargo
Operating Systems

Real-time Facial Emotion Analysis on Jetson Nano

Mon, 01 Dec 2025 00:00:00 +0000

Background

Time: Dec 2025
Context: Main practical assignment for the course “Deep Learning”.

GitHub Repository

https://github.com/Q-Daisy/Ultra-Light-Fast-Generic-Face-Detector-1MB

Highlights

Fine-tuned a lightweight face detection model (CNN with RFB modules) on a custom emotion dataset.
Deployed the model pipeline on Jetson Nano for edge-side real-time inference.

Tech Stack

Python
CNN
RFB Module
Jetson Nano

SUSTech Merch Store Microservices

Wed, 01 Oct 2025 00:00:00 +0000

Background

Time: Oct 2025
Context: Assignment 2 for SUSTech “Distributed and Cloud Computing” (module: Services & API Architectures).
Repository: Tsurumalu/SUSTech-Merch-Store

Goal

Build a microservice-based online merch store for SUSTech-branded products. The system exposes RESTful APIs to clients, isolates database access behind a gRPC service, and streams operational logs to Kafka through a dedicated logging service—all orchestrated with Docker Compose.

Architecture

Three application services run alongside shared infrastructure:

Service	Role	Protocol / Stack
API Service	REST gateway, JWT auth, business orchestration	Flask, OpenAPI 3.0
DB Service	User, product, and order persistence	gRPC + psycopg2
Logging Service	Centralized log ingestion	gRPC → Kafka

Infrastructure containers: PostgreSQL 17 (persistent storage), Zookeeper + Kafka (log channel log-channel).

1
2
3


Client ──HTTP──► API Service ──gRPC──► DB Service ──► PostgreSQL
 │
 └──gRPC──► Logging Service ──► Kafka

REST API

Defined in openapi.yaml and implemented in Flask:

Dialog: GET / greeting endpoint
UserManager: register, deactivate, get user, update profile, login
ProductManager: list products, get product by ID
OrderManager: place order, cancel order, get order details

Protected routes use JWT (Bearer token, HS256, 24-hour expiration). Authentication failures and key operations are forwarded to the logging service.

Data Model

PostgreSQL schema (db-init/init.sql):

products: SUSTech Hoodie, Water Bottle, Notebook (pre-seeded with stock)
users: sid, username, email, password hash
orders: user–product linkage with quantity cap (1–3 per order) and computed total_price

Implementation Highlights

Service decomposition: API layer never talks to PostgreSQL directly; all DB operations go through the gRPC AssistantService.
Connection pooling: DB service uses psycopg2.pool.ThreadedConnectionPool for concurrent gRPC requests.
Structured logging: API service streams LogMessage events (register, login, order, auth failures) to the logging service, which publishes JSON payloads to Kafka via confluent_kafka.
Containerized deployment: compose.yaml wires api-server, db-server, and logging-server with health-checked Kafka/Zookeeper and volume-backed Postgres init scripts.
Dev utilities: Makefile helpers for local psql, DB reset, and Kafka log streaming.

Tech Stack

Python (Flask, gRPC, PyJWT)
PostgreSQL 17
Apache Kafka (Confluent Platform 7.7)
Docker & Docker Compose
OpenAPI 3.0
Protocol Buffers (gRPC service definitions)

Deep Learning for Combinatorial Optimization and EasyCO

Mon, 01 Sep 2025 00:00:00 +0000

背景

时间：Sep 2025 - Present
场景：组合优化问题的深度学习建模与通用求解平台建设。

方法

探索基于图神经网络的数据表示与特征学习机制。
与强化学习框架（例如 REINFORCE）结合，设计训练与推理流程。
参与 EasyCO 开源平台开发，沉淀可复用求解组件。

结果

形成可支持工业应用场景的通用高效求解器开发路径。
在平台层面完成了关键模块贡献与工程化集成。

技术栈

Python
Graph Neural Network
Reinforcement Learning
EasyCO

反思

下一步可强化对大规模实例的可扩展性评估。
需要进一步完善训练稳定性与推理效率的联合优化策略。

Intelligent Sanitation Robot with Deep Learning

Tue, 01 Jul 2025 00:00:00 +0000

Background

Time: Jul 2025
Context: Main practical assignment in the NUS SoC Summer Workshop Program.

GitHub Repository

https://github.com/EpsilonZYJ/NUS-SOC

Highlights

Applied transfer learning on YOLOv7 and designed a parallel multi-object detection pipeline for trash (paper and bottles), cigarettes, violent behavior, and human falls.
Built a crane-style robotic structure on a robot car with servo-controlled grabbing arms and a launching mechanism.
Integrated perception and control with Arduino and Raspberry Pi for end-to-end execution on real hardware.

Poster

Tech Stack

Python
YOLOv7
Arduino
Raspberry Pi
Robotics

Loan Default Prediction with Machine Learning

Sun, 01 Jun 2025 00:00:00 +0000

Background

Time: Jun 2025
Context: Machine learning project for loan risk modeling.

Highlights

Built an end-to-end binary classification pipeline for loan default prediction, including preprocessing, feature selection, training, and evaluation.
Benchmarked Random Forest, Bagging ensemble, and XGBoost under a unified experiment setting.
Tuned XGBoost hyperparameters and selected it as the final model based on accuracy and training efficiency.
Generated final predictions for the test set and documented model limitations with practical improvement directions.

Tech Stack

Python
Scikit-learn
XGBoost
Feature Engineering
Binary Classification

Weibo Personality Analyzer

Fri, 09 May 2025 00:00:00 +0000

Background

Time: May 2026
Context: CS114 course project.
Project title: 微博性格分析师：利用生成式AI洞察社交媒体人格
English title: Weibo Personality Analyzer: Using Generative AI to Gain Insight into Social Media Persona

Goal

The project aims to build a web-based tool called “Weibo Personality Analyzer”. By collecting and analyzing a Weibo user’s public posts, the system infers content themes, writing style, and personality-related tendencies from social media text.

The core role of AI is to use the Qwen large language model to process unstructured Weibo posts, extract linguistic signals, and evaluate whether generative AI can support social media persona analysis in a structured and explainable workflow.

System Workflow

User enters a target Weibo username through the web interface at callitwhatyouwant.cn(now unavailable).
The backend launches a Selenium-based scraper through scraper.py, which simulates browser behavior for more stable data collection.
The system supports collecting around 50 to 300 public posts.
Raw posts are cleaned and converted into a structured Markdown file.
Qwen analyzes the Markdown input and produces a Markdown report.
The Flask backend formats the result and returns it to the web frontend.

AI Analysis

The Qwen LLM is used as the core analysis engine. It receives structured Markdown content and produces analysis across three main dimensions:

Content theme modeling: clusters recurring topics and interest areas.
Writing style analysis: evaluates formality, casualness, emotional tendency, and expression habits.
Personality inference: maps language features to MBTI-inspired dimensions while avoiding unsupported psychological diagnosis.

Web Application

Frontend: HTML, CSS, and JavaScript with responsive design.
Backend: Python Flask routes for request handling, scraper triggering, Qwen API calls, and result formatting.
Deployment: Tencent Cloud ECS server with the domain callitwhatyouwant.cn.

Prompt Engineering

The prompt design went through several iterations:

Version	Prompt Strategy	Result
V1.0	“Analyze these Weibo posts: [text]”	Too general and lacked depth.
V2.0	Asked the model to act as a social media analyst and output theme, style, and personality sections.	Structure improved, but personality analysis was still weak.
V3.0	Added MBTI-based dimensions, negative constraints against unsupported mental-state claims, and required Markdown output.	Reliability and readability improved significantly.

Example prompt direction:

As a social media analyst, analyze the following public Weibo posts from three perspectives: content themes, writing style, and MBTI-inspired personality tendencies. Do not infer private mental health status or make claims unsupported by the text. Output the result in Markdown.

Role of Generative AI

Analysis engine: Qwen performs multi-dimensional text interpretation, including topic, sentiment, style, and persona-related signals.
Development accelerator: generative AI helped draft frontend structure, CSS styling, Flask route templates, and debugging suggestions.
Debugging support: AI-assisted log interpretation helped locate syntax and logic errors during development.
Security suggestions: AI provided reminders such as cookie encryption and safer handling of user data.

Tech Stack

Qwen LLM
Python
Selenium
Flask
HTML/CSS
JavaScript
Markdown parsing
Tencent Cloud ECS

Tools Used

Qwen large language model for post analysis.
Claude 3.7 Sonnet for backend development support.
Bolt.new for frontend design support.

Pipelined RISC-V CPU Design in Verilog

Thu, 01 May 2025 00:00:00 +0000

Background

Time: May 2025
Context: Main practical assignment for the course “Computer Organization”.

GitHub Repository

https://github.com/Diaosi1317092/Computer-Organization-Project

Highlights

Implemented a five-stage pipelined CPU based on the RISC-V 32I instruction set on the EGO1 FPGA board.
Added UART communication and branch prediction support.
Built a Python GUI frontend for real-time visualization of registers and pipeline states.

Tech Stack

Verilog
RISC-V 32I
FPGA (EGO1)
UART
Python GUI

Comment

This is one of my favorite project. Our team spend nearly 100 hours for this course project! Not because of requirements but just feeling interesting to make the theory we learned from class and textbook come true. Every night we gather in a lab room after class and work until 1-3 am, feeling fulfilled and calm. We got 118.5 out of 100 (theoretical maximum score) for this project, thanks for my teammates and my teachers.

Video

Project Posters

Thu, 01 May 2025 00:00:00 +0000

This collection brings together academic posters made for projects in SUSTech and NUS SoC Summer Workshop 2025.

Click on each piece for more details.

Autonomous Driving Video Annotation with VideoLLaMA2

Tue, 01 Apr 2025 00:00:00 +0000

背景

时间：Apr 2025 - Jun 2025
场景：自动驾驶视频理解与关键事件标注。

方法

设计聚焦驾驶关键要素的英文 Prompt，突出自车行为与跟车交互。
基于 VideoLLaMA2 进行微调，提升场景描述准确性。

结果

构建了面向自动驾驶场景的视频标注流程。
模型输出在行为与交互描述方面更加聚焦和稳定。

技术栈

Python
VideoLLaMA2
Prompt Engineering
Fine-tuning

反思

后续可引入更细粒度标签体系，覆盖更多复杂交通状态。
建议补充定量评测指标，便于版本间迭代对比。

openGauss vs. PostgreSQL Performance Benchmark

Sun, 01 Dec 2024 00:00:00 +0000

Background

Time: Dec 2024
Context: Main practical assignment for the course “Principles of Database Systems (H)”.

GitHub Repository

https://github.com/Tsurumalu/DB-Project3

Highlights

Compared openGauss and PostgreSQL with pgbench.
Evaluated query performance, concurrency handling, connection efficiency, data import speed, resource utilization, and security.

Tech Stack

openGauss
PostgreSQL
pgbench
SQL

Range Hood FPGA Control System in Verilog

Sun, 01 Dec 2024 00:00:00 +0000

Background

Time: Dec 2024
Context: Course project for “Digital Logic (H)”.

GitHub Repository

https://github.com/Diaosi1317092/Digital-Logic-Project

Highlights

Designed and implemented an intelligent range hood control system on the EGO1 FPGA board, supporting standby, fan speed adjustment, self-cleaning, smart reminders, lighting, and time query.
Integrated a 4x4 membrane keypad and LCD1602 display to enable gesture-based power control, parameter configuration, and real-time status visualization.
Built a modular Verilog architecture with clock division, debouncing, LCD driving, and finite state machine (FSM) control.
Completed end-to-end hardware logic integration from low-level peripheral drivers to high-level functional orchestration.

Outcome

Project Score: 118 / 100

Tech Stack

Verilog
FPGA (EGO1)
FSM Design
LCD1602
Matrix Keypad

University Activities Posters

Thu, 01 Feb 2024 00:00:00 +0000

This collection brings together posters made for campus activities in SUSTech.

Click on each piece for more details.

Java Swing Match-3 Puzzle Game

Fri, 01 Dec 2023 00:00:00 +0000

Background

Time: Dec 2023
Context: Main practical assignment for the course “Introduction to Computer Programming A (H)”.

GitHub Repository

https://github.com/Tsurumalu/3-Match-Java

Highlights

Designed and implemented a Match-3 puzzle game using Java Swing.
Added multi-level progression, score and move limits, and both manual and auto-play modes.

Tech Stack

Java
Java Swing
OOP

Shortcodes

Fri, 25 Aug 2023 00:00:00 +0000

For more details, check out the documentation.

Bilibili video

Tencent video

YouTube video

Generic video file

Your browser doesn't support HTML5 video. Here is a link to the video instead.

GitLab

Quote

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
― A famous person, The book they wrote

Photo by Codioful on Unsplash

Project Posters

Wed, 01 Mar 2023 00:00:00 +0000

This page collects posters built around more specific communication projects: public topics, historical commemorations, and structured visual messages.

Compared with activity posters, these pieces are more theme-led. The composition gives more weight to symbolism, atmosphere, and visual rhythm while keeping the core message legible at a glance.