When I started at K-Scale, I led the engineering effort on building EMX, our expressive multimodal interface for generating lifelike conversational behavior and facial expressions for our humanoids. With K-Scale AI Day approaching, I took ownership of designing and delivering a context-aware, multi-modal expression engine for K-Bot to debut at the event.

During AI Day, EMX orchestrated a full pipeline of real-time perception and generation across audio, visual, and behavioral modalities. I integrated Google MediaPipe to extract facial and hand landmarks, enabling K-Bot to interpret user gestures and expressions in real time. For affective understanding, I incorporated Hume AI and emotion2vec to derive fine-grained emotional embeddings from audio, allowing K-Bot to infer user intent and modulate expression accordingly. For K-Bot's underlying voice response and tool use, I used OpenAI's Realtime Voice API to generate low-latency audio responses enabling natural, interruption-capable dialogue.

Later, I contributed to kos-sdk, a Python SDK wrapping our Rust-based robot operating system, creating the utilities for hundreds of educators, hobbyists, and inventors to program and interact with America's first open-source humanoids.

Toward the end of my contract, I began exploring RL to train a walking policy for Z-Bot, our miniature bipedal robot. Using ksim, I experimented with policy training pipelines to enable more stable and adaptive locomotion.