Hello, my name is Junrui Zhu. You can call me June Ray, which sounds just like my Chinese name “朱君睿”. I am an undergraduate student at Tsinghua University, majoring in EE.

I have experience researching HCI at Tsinghua, AI-powered data system and video generation model at UC Berkeley.

My research interests lie in Compound AI Systems and Interactive Machine Learning, with a focus on developing AI systems that enable seamless interaction, efficient and high-quality data processing, and adaptive learning. I have been particularly interested in understanding and leveraging foundation models in real-world applications, through the lens of data-driven insights.

I am also an applicant for the 2025 Fall Ph.D. program in Computer Science. Feel free to check out my CV and drop me an email! (juneray2003@gmail.com)

Research

Why did I work on different modalities? Is it a sign that I lack focus and depth in my research interests?

It’s the alchemy of Transformer. For me, after learning the fact that SOTA models, from ChatGPT to Sora, are unified by Transformer, I realized that there are little gap between models across different modalities. I am eager to explore and understand these models in real-world applications.

How did I develop an interest in AI systems?

Following scaling law, increasingly powerful models are emerging, with more parameters, more compute. A single large model may do well in general tasks. However in many specialized tasks scaling offers lower ROI than building a compound AI systems. Say you want a trustworthy output, even with full access to all parameters and training dataset of a model, it’s still challenging. But it’s much easier with a system, because a system inherently offers more control.

Additionally, achieving various performance goals and utilizing AI in budget-constrained tasks is more feasible through system design, because system can be dynamic but tuning models are often costly, particularly with larger models.

Last but not least, I am not saying I am satisfied with current models, I just want to clarify that the system levaraging them are complementary and equally important, and I am optimistic about this approach.