Psyche is building human state models: AI that infers how someone actually feels from physiological signals, the ones that can't be performed. We call it the interpretive layer, the mapping between what the body says and what it means.

“I'm fine. Just a busy week.” The voice is steady. But heart rate variability has been suppressed for thirty-six hours, electrodermal activity is elevated, and sleep has fragmented three nights running. Two channels, one moment, opposite stories. Psyche reads the divergence and says what it means, in language a clinician, a coach, or an AI system can act on.

Most emotion AI trains on actors and posed expressions. We work with live, non-performative signals, supervised by validated ground truth instead of self-report, starting with voice, the one sensor everyone already carries. Models commoditize. No model gains a body. The paired dataset connecting signal to human state becomes more valuable as everything else gets smarter.

Psyche is born out of Carnegie Mellon's Human-Computer Interaction Institute, where our work in wearable sensing and affective computing: pre-print: daily mood states predicted from consumer wearable signals alone, at roughly seventy percent accuracy. Our broader research agenda is laid out in the thesis.

We are working with design partners and clinical collaborators now. If you believe emotional intelligence should not be left out of AI, book a conversation or write to us.