Collaborators: Dr. Jiechao Gao & Prof. Michael D. Lepech - Causal Reasoning & Interpretable ML | LLM, LoRA/DPO, Tsetlin Machines

• Finetuned (LoRA, DPO, etc.) open-source LLMs on the first self-constructed dataset with causal structures, improving causal reasoning and mitigating hallucinations — offering the first empirical evidence linking causality to hallucination.
• Developed an LLM-guided semantic bootstrapping method that distills LLM knowledge into interpretable Tsetlin Machines, achieving BERT-level accuracy without embeddings or inference while preserving clause-level transparency — accepted at Findings of ACL 2026.

Advisor: Prof. Yue Zhao - Trustworthy LLMs & Anomaly Detection | LLM, Anomaly Detection, Causal Reasoning, PyTorch

• Proposed CDCR-SFT, a novel fine-tuning method that internalizes causal DAG-based reasoning, boosting causal reasoning to 95.3% (from 72.9%) on the CLADDER benchmark and reducing hallucinations by 11% on HaluEval — accepted at AAAI 2026.
• Proposed NLP-ADBench, the first comprehensive NLP anomaly detection benchmark, establishing a standardized foundation by showing the strong effectiveness of LLM-based features across 19 methods and 8 datasets — accepted at EMNLP 2025 Findings.
• Developed AD-LLM to evaluate LLMs for NLP anomaly detection, showing strong generalization in zero-shot detection, data augmentation, and model selection — accepted at ACL 2025 Findings.
• Designed a multi-step reasoning framework for LLMs to simulate voter decisions, cutting election prediction error by 77% and exposing bias and overfitting issues in models like GPT-4o and LLaMA 3.1.

Advisor: Dr. Zhaoyuan Su, Prof. Yue Cheng - AutoML & MLSys (Inference) | LLM Inference, Trustworthy AI Systems

• Researched automated methods for generating optimal LLM inference compression strategies for user-specific tasks.

Advisor: Dr. Jiechao Gao, Prof. Brad Campbell - Federated Learning for IoT & Healthcare | Federated Learning, PyTorch, Sparse Networks

• Proposed H-FedSN, a personalized sparse network method for hierarchical federated learning in IoT, reducing communication cost by up to 238× while preserving accuracy via structured masking and Bayesian aggregation.
• Developed FedMetaMed, integrating federated learning and meta-learning to enhance personalized medication strategies across distributed healthcare systems, improving model adaptability and privacy preservation.
• Single-handedly developed Fed-LDR, a federated learning algorithm that employs GCN to enhance spatio-temporal data analysis through local data integration and node-centric optimization.
• FedMetaMed and Fed-LDR were accepted by IEEE BIBM and ICDM respectively, and I gave talks on these research.

Advisor: Dr. Iordanis Fostiropoulos - Distributed ML Execution Framework | MLSys, AutoML, Ray, Optuna, Docker, PyTorch, Pytest

• Contributed to an open-source project for distributed AutoML execution used by 30+ researchers at USC.
• Developed Ablator, a deep learning framework for horizontal scaling of ablation experiments, enabling large-scale experiments with Ray and hyperparameter tuning with Optuna, across 50 pull requests.
• Constructed CI pipelines with GitHub Actions, enhancing code quality and package management; authored docs and unit tests with Pytest, increasing test coverage to 95%+.
• Created and published python-rclone on PyPI, simplifying cross-platform use of RClone without pre-installation and with automatic binary selection.

Advisor: Prof. Guoquan Wu - Automated Testing Platform Development | Docker, Node.js, JSON, Vue.js, RobotFramework

• Contributed to the R&D of a web-based automated testing tool using Record and Playback technology, significantly enhancing test case management by enabling streamlined recording, editing, execution, analysis, and result generation. This implementation boosted end-to-end testing efficiency by 300% and saved over 15 hours per week.
• Independently developed a script parser using Node.js that converts user actions recorded in JSON format into executable Robot Framework and Selenium scripts, enabling the replay and repeated execution of these user actions.
• Single-handedly created innovative UI components using Vue.js and AceEditor, orchestrated the optimal containerization of the program with Docker, and automated the DevOps pipeline to maximize development efficiency.

“Hexagon Adventure” 2D Game | Unity, C#, Firebase, Python, Data Analytics

• Implemented automated data collection in Unity using C# to gather critical player behavior and game state data, stored in Firebase Realtime Database to support decision-making.
• Analyzed complex game data using Python, focusing on key metrics like player progression, health point loss, and power-up usage, contributing to a 40% increase in player engagement.
• Drove game design enhancements and level mechanics refinements through data analysis and playtesting sessions, aligning the game with audience preferences.

SaaS Platform Development | Kubernetes, Docker, Go, CI/CD, CRD, Operator-SDK, Helm3, Prometheus, Grafana, ELK

• Developed “RocketMQ as a Service”, mirroring “RabbitMQ as a Service” from AWS Marketplace: fully managed RocketMQ clusters with robust support, easy monitoring, convenient scaling, and seamless updates.
• Utilized Operator SDK and Custom Resource Definitions (CRD) to construct a Kubernetes-based RocketMQ Operator, automating lifecycle management and minimizing manual intervention — leading to 50% faster RocketMQ service creation, automated error recovery, and improved reliability.
• Employed Helm3 to package components into Helm charts, simplifying Kubernetes deployment; integrated Prometheus and Grafana for real-time monitoring of key metrics and node health.
• Automated key workflows including unit tests, container image builds, and Helm3 Chart updates via GitLab CI/CD, speeding up development processes.

Text Translation Platform Development | Java, SpringBoot, Spring, Java Persistence API, Maven, Nginx, MySQL, Git

• Developed an AI document translation system with Java/Spring/Maven; independently responsible for the parsing and translation of PDF, XML, and EML documents, attracting 30,000 MAUs.
• Used Nginx for reverse proxy and load balancing, and managed version control with Git.

Recruitment Data Mining System | Python, Spring Boot, ReactJS, NodeJS, MongoDB, MySQL, Docker

• Led the development of a data mining system to gather recruitment data and conduct multi-dimensional statistical analysis; earned “National Level Innovative Excellence Project” recognition and was adopted by the school.
• Utilized web scraping to extract 10 million+ recruitment records from major job sites with Python, lxml, and XPath; used Bloom Filter for deduplication and Docker with CI/CD for project automation.
• Employed Spring Boot, ReactJS, NodeJS, AntV, and SQL to build automated reports, dashboards, and data visualizations; automated ETL processes with numpy and pandas, storing data in MongoDB and MySQL.