Lianrui Geng

Lianrui Geng

Computer Science Master Student in UNC

Biography

Lianrui (Bruce) Geng

I am a graduate student pursuing a Master of Science in Computer Science at the University of North Carolina at Chapel Hill (UNC), and a Graduate Research Assistant at ZDlab (advisor: Prof. Zhun Deng). My work sits at the intersection of Large Language Models, Human–AI Collaboration, and multi-agent systems — specifically, how to make LLMs more calibrated, honest, and useful when they collaborate with humans on high-stakes tasks.

I came to LLMs from a strong full-stack and systems background. At Utah, I led a Top-1 capstone — an open-sourced legal-document analysis platform that has earned 80+ GitHub stars — and shipped projects in Unreal Engine 5, GPU ray tracing, and interactive C++/Qt teaching tools. I treat that foundation as a feature, not a bug: it means I can take a research idea and ship a usable product around it.

🎯 I’m actively looking for Summer 2026 internships in Applied AI / LLM engineering or ML research engineering. CPT-authorized, US-relocatable. Reach out if you’re hiring.

🎓 Education

  • University of North Carolina at Chapel Hill (Chapel Hill, NC)
    • Master of Science in Computer Science — GPA 4.0
    • Aug 2025 – May 2027 (expected)
    • Graduate Research Assistant, ZDlab — advisor Prof. Zhun Deng
    • Research focus: Human–AI Collaboration, multi-agent LLM systems, prompt optimization
  • University of Utah (Salt Lake City, UT)
    • Bachelor of Science in Computer Science, Minor in Mathematics
    • Graduated: Dec 2024 — GPA 3.6/4.0
    • Honors: Utah Global Scholarship (4 academic years, $6,000), Dean’s List (5 semesters)
    • Achievement: “Staff Choice” Award + Top-1 Capstone Project, Fall 2024

🔬 Research Experience

Graduate Research AssistantZDlab, UNC–Chapel Hill   Nov 2025 – Present

Building two end-to-end LLM research pipelines from scratch — covering data, training, evaluation, and multi-GPU infrastructure.

  • CollectiveLLM — LLM survey simulation & DST-guided interviewer fine-tuning
    • Built a multi-turn “interviewer ↔ simulated respondent” rollout system with batched generation, Monte-Carlo question search, and shared-engine vLLM inference across 8 GPUs.
    • Trained a question-aware Dialogue State Tracking (DST) model (DistilBERT backbone with question-conditioned head) to recover respondent stances; used DST scores both to filter SFT data and to construct preference pairs for DPO.
    • Implemented the full interviewer fine-tuning loop with HuggingFace + PEFT/LoRA SFT and TRL-based DPO, then regenerated higher-quality rollouts to form a self-improvement cycle.
    • Reproduced the pipeline on 4 survey datasets (CES, WVS, Polish Participatory Budgeting, OpinionQA).
  • HAC — Human–AI Collaboration framework for medical MCQA
    • Designed a selective-prediction pipeline that lets an LLM answer when confident and defer to a (simulated) human otherwise, optimizing the custom HAC loss L(τ) = (1 − sel_acc) · coverage + λ · human_rate via threshold search.
    • Implemented a unified confidence-extraction stack — Maximum Softmax Probability, margin, negative entropy, verbalized self-evaluation, and explicit “defer” prompts — benchmarked with ECE, AUROC, and risk-coverage curves.
    • Ran systematic ablations across 12 prompt templates and 3 framework variants on Qwen2.5-7B, Llama-3-8B, and BioMistral-7B over MedMCQA, MedQA, and PubMedQA.
    • Built the multi-GPU evaluation infrastructure (torchrun DDP, model-agnostic local-HF / vLLM / OpenAI wrappers, reusable plotting scripts).

💼 Professional Experience

Game Development Engineer Intern | Aureum Gale Game (Los Angeles, CA) May 2024 – Aug 2024

  • Developed a fully functional 3D game prototype using Unreal Engine 5, used for critical team presentations.
  • Integrated 20+ unique game assets, enhancing user engagement and playability through close collaboration with the design team.
  • Managed version control via Git, achieving a 30% reduction in code conflicts and streamlining the team workflow.

Remote Web Developer | Heilongjiang Guanjie Biotechnology Co., Ltd (Harbin, China) Apr 2024 – Aug 2024

  • Designed and maintained 3 responsive websites with HTML/CSS and JavaScript, lifting user retention by 30%.
  • Boosted application performance by 40% via interactive dashboards and mobile support optimization.
  • Coordinated weekly sprint cycles remotely, improving development efficiency by 25%.

Computer Service Support Administrator | Jilin Baiqi Pharmaceutical Co., Ltd (Jilin, China) May 2023 – Aug 2023

  • Reduced system downtime by 25% through proactive hardware/software maintenance.
  • Upgraded critical network infrastructure, increasing communication reliability by 30%.

💻 Selected Projects

Wrap — Intelligent Legal Document Assistant (Capstone, Top-1 Fall 2024) Full-Stack Engineer & Project Lead · React, Node.js, MongoDB, AWS, Python/PyTorch, Chrome Extension Aug 2024 – Dec 2025 · wrapcapstone.com · github.com/LianruiBruce/wrap — 80+ ⭐ / 9 forks / 192 commits / 4 contributors

  • Top-1 Capstone + Staff Choice Award. Led design and full-stack development of a legal-document summarization platform deployed on AWS with NGINX-secured communication.
  • Built a Chrome extension that auto-detects Terms / Privacy Policies / Contract Agreements on the page and pipes them through a Python/PyTorch + Flask NLP backend (CUDA-accelerated) plus the Claude API for summarization and report generation — lifting document-processing efficiency by 40%.
  • Connected extension and web app via JWT auth so users can upload documents, generate PDF reports, and access historical and current summaries from a single account.
  • Open-sourced under MIT; production deployment served real users at wrapcapstone.com.

High-Performance Ray Tracing with DDA Algorithm Computer Graphics Engineer · C++, OpenGL, GPU Programming · Apr 2024 – May 2024

  • Implemented the Digital Differential Analyzer (DDA) line algorithm on the GPU, bypassing CPU floating-point operations.
  • Achieved a 50% reduction in ray-tracing computation time.
  • Simulated complex visual effects, including dynamic objects and mirror rendering, resolving oversampling issues to ensure visual fidelity.

Machine Learning Educational Application Software Engineer · C++, Qt, Box2D · Nov 2023 – Dec 2023

  • Architected an MVC-based educational tool to demonstrate Linear Regression concepts.
  • Integrated Box2D for physics simulations and polynomial curve generation, visualizing mathematical principles through interactive gameplay.

🛠 Technical Proficiency

  • Languages: Python, C/C++, C#, Java, JavaScript, TypeScript, SQL, MATLAB, HTML/CSS
  • AI / ML: PyTorch, HuggingFace Transformers, PEFT / LoRA, TRL (SFT, DPO), vLLM, torchrun DDP, scikit-learn, prompt engineering, evaluation (ECE, AUROC, risk-coverage curves)
  • Backend / Web: Node.js, Express.js, Django, Flask, React, Material-UI, Axios, Mongoose, REST APIs, Socket.IO
  • Data / Infra: MongoDB, MySQL, Docker, AWS, NGINX, Git, multi-GPU Linux clusters (CUDA)
  • Specializations: LLM training and evaluation, Human–AI Collaboration, full-stack delivery, computer graphics, Unreal Engine 5 (Blueprints)

Feel free to connect via LinkedIn or explore my work on GitHub.