Synheart's infrastructure is live. On-device inference is free, forever.
Synheart's infrastructure is live. On-device inference is free, forever.
SYNHEART RESEARCH
Synheart researches how physiological signals, movement, and digital behavior can be modeled with machine learning and behavioral science to infer human state in real time. Every insight becomes part of the Human State Interface (HSI), the infrastructure behind our products and SDKs.
RESEARCH AREAS
What can the body tell us about changing human state?
How can physiology, behavior, motion, and context be combined to infer human state?
What does sustained attention in HRV and motion data look like?
Can micro-motion and timing patterns reveal behavioral state?
How can models learn each person’s baseline and track state changes over time?
Can human-state intelligence stay accurate and personalized while raw data remains on-device?
CASE STUDIES
Each case study follows one research question from hypothesis to finding, and where relevant, into what shipped. 2 active studies, more added as they progress.
A Research Tool From Synheart Uses Cognitive Games To Collect Multimodal Data: Physiological Signals, Behavioral Patterns, And Interactions. It’s Open To Researchers And Developers.
An active research program investigating how real-time human state signals, stress and focus among them, can be integrated into language model inference to produce responses adapted to a person’s state at the moment of interaction.
Publication expected, follow for updatesCan a language model with real-time human state context produce meaningfully better, more contextually appropriate responses than one operating on text alone?
OPEN SCIENCE
Validates the Focus Performance Score from the Pulse Focus Stroop game as a reliable, neurally grounded measure of attentional control.
Models cognitive capacity as two physiological states, mental effort and stress, estimated from wearable cardiac and electrodermal signals.
Defines a privacy-first, on-device standard for human state: the Human State Vector to represent it and the Human State Interface to exchange it.
Turns taps, scrolling, idle time and app switching into numerical behavior metrics, without collecting any content or personal identifiers.
Compares classical, deep learning and transformer models for on-device emotion recognition from wrist PPG, and deploys a compact ONNX model.
HOW WE WORK
Inference runs entirely on-device. Raw biosignals never leave the user’s phone or wearable unless the user explicitly consents. Human State Intelligence should never come at the cost of privacy.
Our work doesn’t stay in papers. We design research with deployment in mind, prioritizing robustness, interpretability, and real-world constraints from day one. Every insight becomes part of HSI.
We publish our research, open-source our models, and release our datasets publicly. The biosignal research community should be able to inspect, replicate, and extend our work freely.
WHO BUILDS ON SYNHEART
The PAMAP2 benchmark (Reiss, 2012): 18 physical activities recorded from 9 people wearing three motion sensors and a heart-rate monitor.
ExtraTrees models that tell baseline from stress using heart rate variability features, built to run on the device with ONNX Runtime.
Google’s Gemma 3 1B Instruct in 4-bit GGUF (Q4_K_M), packaged so Syni, the companion in Life by Synheart, can run it on the device.
AND WHAT’S BUILT ON IT
Synheart strengthens research via partnerships and collaborations with universities and labs. We focus on privacy-first, human-centered AI. All collaborations, including early agreements, are listed here.
View all partnersWhere our work is shared: open datasets, code, papers, and the wearables we support.
Researchers who advise our work in physiology, human-computer interaction, and machine learning.
Research Advisor
Sammy Assefa
HCI
Alemitu Bezabih
MILA / MHI
Alexis Nolin-Lapalme
Universities we work with on human-state research.
New papers, datasets, and case studies as they publish.
No spam, unsubscribe anytime
