About
I’ve always been drawn to the space where art and machine learning meet. Now I’m building Kivo—a multimodal agent that understands graphic tools and learns to design through action.
Can a model learn design inside the same tools—and under the same constraints— as a person?
Since 2020, I’ve worked on generative models—starting with normalizing flows for diverse super-resolution, then moving into transformer-based image generation. I first-authored Noise Conditional Flow Model at CVPRW 2021 and was a first author of Styleformer (equal contribution) at CVPR 2022.
In 2023, I founded OptimizerAI to turn original generative-model research into creative tools. I first built a flow-matching sound foundation model from scratch; today, OptimizerAI is where I’m building Kivo. The company was backed by a16z Speedrun.
Now I’m building graphic environments where agents can see the canvas, use tools, and receive feedback. I’m interested in whether reinforcement learning can teach a model not just to produce an image, but to make design decisions more like a person.
Current thread: graphic environments × reinforcement learning.