Zekai Huang.

Lately, I’ve become more interested in why a method works—and under what conditions. When studying agents, multimodal models or post-training, I keep asking: what goes in and comes out, what problem does each module solve, and would it still work in my setting? Understanding the mechanism helps me decide whether to change the model or first improve the data, context and processing flow.

Much of my recent work has involved reading source code, orchestrating implementations and adapting open-source projects to business needs. These experiences are gradually helping me understand how complex systems grow from small modules, integrate into business workflows and deliver sustained value. My internship has also made me think about how to go deeper into specific problems within multimodal video algorithm modules. I think a useful approach is to start with a concrete problem, use realistic examples and controlled comparisons to understand what improves and where the method fails, then build on a solid small experiment toward a complete system.

I enjoy looking for connections across fields. Could representations and indexing from retrieval and recommendation help a long-running agent recover information relevant to its current task? Could video examples that preserve evidence and corrections become a starting point for post-training smaller models to understand actions and changes in state? These remain open questions that I want to test through practice.

AI helps me explore more broadly and implement ideas faster, but goals, tradeoffs and judgment remain my responsibility. Through experiments, feedback and refactoring, I hope to develop a stronger sense of what matters.