Yunkai (James) Zhan

zhanyunkai603@gmail.com  ·  +1 (213)-716-2016  ·  Los Angeles, CA
LinkedIn: https://linkedin.com/in/yunkai-zhan
Email: mailto:zhanyunkai603@gmail.com

EDUCATION
=========

University of Southern California, Viterbi School of Engineering   (Aug 2023 - May 2027)
Bachelor of Science, Computer Science | Applied Math (double major)

EXPERIENCE
==========

Currents AI — Machine Learning Engineer (Internship) — Social World Model, Post Training   (Oct 2025 - Present)
  - Building an LLM-based Celebrity Simulator to simulate the top 100 political, business, and finance figures' tone, knowledge, belief system, and decision making with a post-training pipeline and agentic framework.
  - Establishing a post-training pipeline with multi-source data collection and processing, Supervised Fine-Tuning (SFT) on Qwen, and reinforcement learning with a generator-discriminator framework.
  - Participating in the company's web agent extension launch, including test script writing and marketing planning.

Alibaba Cloud — Software Engineer (Internship) — Knowledge Graph, RAG   (Jun 2024 - Aug 2024)
  - Pioneered the industry's first RAG multi-hop question set auto-generator for the Chinese finance sector; evaluated and improved the model's retrieval accuracy by 20%.
  - Developed a Python pipeline based on Knowledge Graph (Neo4j), vector embeddings, and prompt engineering to auto-generate more challenging questions for the RAG system.
  - Applied the pipeline to Alibaba's finance AI assistant Qwen Dianjin's internal RAG evaluation, improving product satisfaction among 30+ leading Chinese financial companies.

PROJECTS
========

Multimodal LLM Adaptive Thinking — Academic Research — MLLM Post-Training   (Feb 2026 - Present)
  - Post-trained multimodal LLMs (Qwen3-4B-VL) across diverse vision-language tasks under both thinking and non-thinking modes; identified a bidirectional performance gap where thinking mode outperforms by ~15% on geometry and math but underperforms by 5–10% on VQA and counting tasks.
  - Designing GRPO-based training with mode-conditioned reward shaping to unify reasoning capabilities, enabling a single model to adaptively leverage extended or compressed thinking without sacrificing performance in either regime.

Research Assistant — USC HUMANS Lab — Academic Research — Social Simulation, Agent-Based Modeling, RL   (May 2025 - Present)
  - Contributed to a WWW 2026 submission on simulating LLM-driven influence operations (IO) using multi-agent generative modeling in social media environments.
  - Helped implement a Generative Agent-Based Model (GABM) of 40 organic and 10 IO agents using AutoGen and Llama 3.3-70B, exploring coordination under varying operational awareness levels.
  - Researching the application of reinforcement learning to multi-agent social media simulation to improve simulated online marketing engagement rate and purchasing intent uplift.

Research Assistant — USC Media Communication Lab — Academic Research — RL, Green Learning   (May 2025 - Present)
  - Conducting research with Prof. Jay Kuo on the intersection of Reinforcement Learning and Green Learning — an explainable, energy-efficient alternative to conventional deep learning pipelines.
  - Surveyed and reproduced 10+ papers on main approaches in Green Learning on standard vision benchmarks, focused on feature engineering (variants of PCA, Sparse Coding) and XGBoost.
  - Researching the application of the Green Learning pipeline on Policy Gradient as an alternative to Deep Policy Networks with high interpretability, efficiency, and superior performance on small datasets.

LLM-Powered Learning Assistant — Project — Next.js, TypeScript, MongoDB, RAG, LLM   (May 2025 - Aug 2025)
  - Engineered a production-ready Retrieval-Augmented Generation (RAG) pipeline using FastAPI, embedding course materials with Voyage-3 into an L2-normalized FAISS vector index for high-speed semantic search.
  - Orchestrated a multi-model LLM strategy, leveraging Claude 3.5 Haiku for efficient document summarization and Claude 3.7 Sonnet for context-aware answers, achieving 99% concept accuracy.
  - Built a full-stack interface featuring an interactive document viewer and persistent chat backed by MongoDB for scalable storage of large PDFs.

SKILLS
======

Coding: Python (NumPy, Pandas, Scikit-Learn, PyTorch), C++, Java, MATLAB
Web Development: React, Node.js, HTML, CSS, TypeScript, JavaScript, Spring Boot, MySQL, Git, Docker
ML / AI: LLM Post-Training (SFT, RLHF, GRPO), RAG, Knowledge Graphs (Neo4j), Vector Embeddings (FAISS), Multi-Agent Systems (AutoGen), Green Learning
Relevant Coursework: Algorithms, Data Structures, Software Engineering, Machine Learning, Linear Algebra, Probability and Statistics, Data Science

cv.ha7ch.com/james