Unsubtle

Non-verbal communication is full of signals that many people rely on without realizing it — a crossed arm, a shift in posture, the subtle tension in someone's shoulders. For individuals who struggle with verbal communication, or those supporting them, these cues can be the primary channel. Unsubtle makes them visible.

The project originated from personal experiences working with students in special-needs classrooms — a recognition that the challenge isn't a lack of expression, but a lack of tools to bridge the gap between what someone communicates and what others perceive. Unsubtle uses computer vision to detect body language in real time and renders it as accessible visual feedback directly in a Meta Quest 3 headset.

"We built something with the potential to help people feel more understood — especially those whose communication needs are often overlooked."

Interface & Interaction Design

My work focused on the real-time visualization layer inside the headset: designing how detected body language is represented on-screen in a way that's legible, non-stigmatizing, and genuinely useful in the moment. The system renders a user's skeleton and dynamically changes its color to reflect recognized poses — crossed arms signaling defensiveness, open posture signaling engagement — translating invisible cues into clear, immediate visual language.

I also designed the microgesture interaction system, which lets users send discreet, low-effort signals to others — an emoji, a short phrase like "I need personal space" — without requiring speech. These outputs sync across colocated XR devices via a shared networking layer, which I helped implement and stabilize to ensure UI state and display data remained consistent for all participants simultaneously.

Research & Grounding

Throughout development I conducted user research with educators and caregivers to make sure the interaction model reflected real-world needs rather than assumptions. A recurring finding: the tool had to feel supportive, not diagnostic. Body language should be treated as a cue to understanding, never a verdict. That principle shaped every design decision, from the color system to the phrasing of microgesture outputs.

The backend runs a fine-tuned pose estimation model on a FastAPI server, receiving image data from the headset via POST requests and returning detected positions for real-time rendering — a hybrid architecture we pivoted to after finding on-device inference too unstable for the live use case.