metadata
title: README
emoji: ⌚
colorFrom: green
colorTo: gray
sdk: static
pinned: false
Somatic AI · 映身智能
Human Body Tokenizer for AGI
Making the human body readable for AI.
We are building Soma, a foundation model for continuous physiological signals from wrist wearables. During training, sports video, audio, metabolic-chamber and lab measurements label the wristband's raw streams: PPG, accelerometer and skin temperature. At inference, the model needs only the wristband.
Our models will be published here.
First release: Soma-ACC (coming soon)
A cross-device wrist accelerometer model. Every 30 seconds, it tells what a person is doing and whether they are asleep, from wrist motion alone. No camera, no microphone.
| Input | Raw 3-axis wrist accelerometer data from any device. The model does not need to know which device. |
| Output | Every 30 seconds: an activity label scored against text descriptions of activities, plus a sleep probability. |
| Training data | Public datasets only. Every dataset is credited in the model card. |
| Evaluation | On people held out from training, with the device treated as unknown. |
Principles
- Public data, credited. Every dataset we train on is named and credited in the model card.
- Held-out people, unknown devices. We report results on people the model never saw, with device identity hidden at test time.
- Models produce the numbers; language models phrase them. LLMs never generate physiological values.
- Wellness and research, not medicine. Our models are not medical devices and are not intended to diagnose, treat or prevent any disease.
Related
- SomaLoop: a screenless research wristband that streams raw PPG and accelerometer data.
Have reference measurements, such as sports video, metabolic-chamber data or lab tests? Let's pair them with continuous wrist signals: contact@somaticai.io