At KAIST Interaction Lab, I designed reusable components for Maru, an LLM-driven generative UI system. They help preserve and retrieve users’ information architecture, layout, and interaction preferences across conversations.
At Big Data for Smart Society: GATE Institute, I’m investigating reasoning and planning failures in LLM agents, including how those failures affect long-horizon task completion. I’m also developing a taxonomy of agent planning approaches and failure modes.
At SignVrse, I designed experiments on hierarchical representations in a MoMask RVQ-VAE for sign-language motion generation. My analysis found that the first residual codebook recovered approximately 71% of the reconstruction gap, and the first two recovered 91–92% of full six-codebook quality.
I built UbuntuTech, a Shona writing assistant combining linguistic rules with an LLM correction layer, and developed a segmentation pipeline for Ghanaian Sign Language videos that outperformed seven baselines by up to 33% in Adjusted Rand Index. My work also includes scalable market-data pipelines at Princeton University and automated testing at Econet Wireless Zimbabwe.

