I built an AI chat platform called Segment powered by large language models like GPT, Gemini, and Llama, customized with specific instructions. The challenge? Making all these different models compatible with each other. To solve this, I developed a creative architecture: a generic chat object format that orchestrates communication across APIs. This allows users to seamlessly switch from a conversation with Gemini to GPT mid-session, preserving the entire history within the generic State. It essentially functions as a lightweight custom SDK for AI models.

Curiosity drove me to explore Three.js, a powerful library for rendering 3D graphics directly in the browser, from simple geometries to complex environments. However, production-ready 3D web graphics are expensive and resource-intensive. I initially built a 3D room for my portfolio, but the initial setup ran into severe performance issues. To fix it, I researched optimization strategies: compressing `.glb` files using Draco, leveraging conditional `frameloop` execution to render frames only during active user interaction, generating optimized React components from raw assets, and lowering segment counts. As a result, I cut down the 3D room's loading time from 5.7 seconds to under 2 seconds, achieving a near-unnoticeable impact on visual fidelity. You have probably already interacted with it if you are viewing this page on a desktop.