ResearchGateResearch and Engineering Projects
A dedicated overview of ongoing and early-stage work.
Ongoing
Current FocusZhiYuDao (CancerDAO) — AI Health Companion for Cancer Patients
As project lead, I built a WeChat mini-program that turns patients' scattered medical records into a structured, longitudinal health archive — AI report parsing (OCR + LLM extraction across 8+ document types, checked against current oncology guidelines), indicator-trend tracking, and an AI consultation assistant.
Architected on Tencent CloudBase serverless (~126 cloud functions, MongoDB), with Coze-orchestrated LLM workflows and blockchain (SBT) data-ownership and consent auditing.
Shipped and in acceptance testing; now extending to native iOS/Android (uni-app) and a self-hosted LLM backend.
ODBB — Oncology Decision Boundary Benchmark
First-author study (under review at npj Digital Medicine) on whether frontier LLMs share systematic blind spots in real oncology decisions — not factual recall, but choosing guideline pathways and committing under uncertainty. I built ODBB: 2,005 decision points across NCCN guidelines and colorectal-cancer cases, scored by a fully deterministic engine and validated by two oncologists.
Across nine frontier models (GPT-5.5, Claude, Gemini, Qwen, DeepSeek, GLM…), pooling all nine as a single "super-model" still left 42% of items answered correctly by none — a collective capability boundary rooted in cross-guideline meta-judgment, not knowledge. Benchmark and scorer openly released.
Early Projects
FoundationsCancerDao - Blockchain Ownership Workflow
Completed patient-centric access control with SBT identity, token-level ACL, and on-chain audit logs.
Delivered full workflow: key management, consent signing, minting, and admin operations via a Node/Express + PostgreSQL gateway.
Smart Contract Fraud Detection and Obfuscation Analysis
During my visit to Singapore Management University under Prof. Yue Duan and Prof. Kai Li, I designed Obs_tool, a transfer-centric, bytecode-level analysis framework that systematically uncovers and quantifies obfuscation in Ethereum contracts. Building a taxonomy of seven key obfuscation features, I applied Z-score ranking to analyze 1.03 million contracts, isolating over 3,000 highly obfuscated instances.
This study revealed two novel scam types, three high-risk contract categories, and MEV bots employing assembly injection, dead code, and deep function splitting. I further demonstrated that obfuscation significantly increases both the scale of financial damage and evasion time, and showed how the state-of-the-art Ponzi detector SourceP suffers an accuracy drop from about 80 percent on clean samples to about 12 percent on obfuscated ones.
AI and Blockchain Fraud Detection
As part of my research at Zhejiang University with Prof. Haitao Xu, I proposed the ETH-GBERT model, which combines Graph Convolutional Networks (GCN) and BERT to detect fraudulent activities in blockchain transactions. The project involved designing a multimodal learning framework, creating a data pipeline for transaction analysis, and achieving significant improvements over baseline models. PDF
Acoustic Beam-Tracing Software GPU Acceleration
In this project, I accelerated an industrial acoustic simulation software using CUDA. By optimizing memory access patterns, parallel execution strategies, and hardware resource utilization, I achieved a nearly 800x performance improvement. This work also involved rewriting and restructuring legacy code for GPU compatibility.
GPUPixel Project
During my internship at iFlytek AI Research Institute, I worked on the GPUPixel rendering framework. I implemented features such as virtual server-based rendering, enhanced image output functionality, and EGL adaptation to improve the system's performance and future compatibility.
Acoustic Beam-Tracing Software Refactoring
I refactored a 7,000-line Fortran acoustic simulation software into C++. This work focused on ensuring functional consistency, improving code maintainability, and preparing the system for subsequent optimizations and extensions.
E-commerce Recommendation System
For my undergraduate thesis, I designed and implemented a recommendation system for an e-commerce platform. The system analyzed user behavior, search history, and purchasing patterns to deliver personalized product recommendations.
GPS and Sensor Technology for Epidemic Prevention
This project focused on integrating GPS and sensor data to improve epidemic prevention measures. I led the design and development of the system, including architecture planning, frontend interface creation, and backend interaction testing. The project also produced a research report summarizing the theoretical and practical findings.