Make the human body
readable for AI.
Human Body Tokenizer for AGI
From raw signal to body states AI can call. The SomaLoop band records raw signal continuously, video, audio, metabolic and lab data label it, and Soma turns it into states, forecasts and simulations with confidence attached.
AI, wearables and longevity have arrived at the same place.
AI has lowered the bar for using health data
More than 300 million people a week ask ChatGPT health questions, and every health assistant is reading personal data. Producing data, and connecting it to AI, matters more than ever.
Wearables are the most continuous physiological data
Worn every day, they are the most frequently used health product. Raw PPG arrives 50 times a second; a blood test comes once or twice a year, a genome once in a lifetime.
Longevity and the biohacker community are growing
The biohacker community on Reddit is about 900,000 people and grew nearly 40% in a year. More people are willing to invest in living well for longer.
The data is locked away, and models and infrastructure have not caught up.
Wearables are hard to connect to AI, and raw data is mostly closed
Official APIs from the major brands mostly expose summary metrics and no raw PPG. Communities reverse-engineer Bluetooth protocols just to get their own data back.
No foundation models or AI infrastructure for continuous physiology
Big-company models stay in papers and learn from minute- or hour-level aggregates. Synchronised capture and alignment, reference-measurement labels, evaluation and calibration, and interfaces agents can call are all left to developers to build themselves.
One number cannot read a person.
Nothing between two points
Four points a year and three hundred unobserved days. Change has a before and an after, but no date and no cause.
Small changes hide in noise
Every single measurement carries technical noise, and modest real change is often indistinguishable from it. Claims like "biological age dropped five years" deserve caution.
Averages swallow the truth
Same metabolism, same weeks: one HbA1c reading of 5.0%, "perfect"; two weeks of continuous glucose, GMI 5.8%, every post-meal excursion visible. One method saw it. The other did not.
Device, model and harness: three layers from raw signal to body states AI can call.
Raw signal
Continuous, traceable to the waveform
Multimodal labels
Video, audio, metabolic, lab, diet
Foundation model
One backbone, many tasks
Harness
Interfaces, evaluation, confidence, audit
Applications
Institutions, developers, agents
A wearable that captures raw signal accurately
Training needs comparable, continuous signal traceable to the raw waveform, not each brand's private score. Worn for long periods, with exportable raw data and one consistent capture protocol.
A foundation model trained on multimodal data
Video, audio, metabolic, blood and omics tests, diet and supplements label the continuous physiological stream. One model assesses current state, forecasts change and simulates the effect of interventions.
Infrastructure and application systems
Around the model sits a layer of engineering: ingestion, alignment and labelling, evaluation and calibration, audit and call interfaces. Institutions, developers and agents can call it directly, and every result is traceable.
A band, an app, a developer toolkit and a model.
SomaLoop band
Research-grade data in a consumer band. Exportable raw PPG and acceleration, worn on the wrist or upper arm.
Hardware →SomaLoop for BioHacker
A band that opens its raw data and an AI companion that knows you. iOS closed beta under way; Android in development.
App →SomaLoop SDK and cloud
Raw signal in the hands of developers and institutions: dual-platform SDK in public beta; data sync, gap filling and analytics running in a test environment.
Developers →Soma
Multimodal labels to assess state, forecast change and simulate interventions. First model Soma-ACC is coming soon as open source.
Model →Every partnership adds a capability to the same model.
Labelling the band with sports video
With a team at Central South University we built a synchronised capture system: video recognises pose, shots and rallies, aligned in time to the band's accelerometer to produce event-level labels. The alignment and segmentation tool, SomaLoopMotion, is public on GitHub. Next: the first professional motion-recognition model, so courts without cameras can recognise shots and rallies from the band alone.
Labelling the band with a setlist timeline
The concert's setlist timeline labels heart rate and movement from every band in the venue, showing which song and which passage moved the crowd. Sessions with band signal paired to the setlist already exist. Conclusions are drawn on the crowd only, never on individuals.
More partnerships running or under way
Research services are running with teams including the School of Artificial Intelligence at Central South University, and devices are in application trials at two national-level institutions. Several leading longevity clinics in China and abroad are moving toward continuous monitoring between tests. A 24-hour standardised metabolic-chamber protocol, with minute-by-minute energy expenditure as the label, is about to start.
Four kinds of partners, combining device, data and model services as needed.
Research institutions
- For
- Sports science, rehabilitation, sleep and metabolism teams that need continuous physiology.
- We provide
- Research bands with raw data export, a data platform, custom capture protocols and dedicated models.
- How
- Joint research, collaborative projects, device and data services.
Longevity clinics and health-management providers
- For
- Providers already running regular tests for their members.
- We provide
- Continuous monitoring between tests, state and trend with confidence attached; on-premise deployment available.
- How
- Per-person-per-day service; tests paired with band data and labelled together.
Developers and ecosystem partners
- For
- Teams building health apps, AI coaches and agents, sports equipment and venues.
- We provide
- iOS and Android SDKs, APIs and /llms.txt; replay recorded real data to build without a band; MCP to follow.
- How
- SDK and API integration; co-developed dedicated models such as motion recognition.
Data and labelling partners
- For
- Teams that hold reference measurements: sports video, metabolic chambers, functional-medicine tests, questionnaires and expert assessment.
- We provide
- Synchronised capture protocols, alignment and labelling tools, and the continuous-signal side of the data.
- How
- Joint capture and labelling, joint model validation.
From a tokenizer for the body to a world model of the body.
Read the state
How is the body now?
Passive recording and event labels: badminton, concerts, daily sleep, recovery and load scores, plus everyday activity recognised every 30 seconds.
Perturbation and response
After I do this, where does the body go?
Controlled interventions: standard meals and cycling in the metabolic chamber, standardised mini-experiments in the app.
A world model of the body
If I do this, what happens to me?
Prospective validation: randomised self-experiments and long-term follow-up.
Perturbations include exercise, diet, schedule and medication. Medication is a model input from day one to correct the baseline; the model accounts for medication but never recommends it.
In the age of AGI, every AI that serves a person, agent or robot, will need to understand the person it serves.
Make the human body readable for AI.
Research institutions, longevity clinics, developers and data partners are welcome.