Synheart Synheart synheart.ai ↗

Synheart,
to you.

Tell us how your company uses AI. We'll read it ourselves and write back with a short study of where Synheart fits, at an address of your own.

synheart.to/

Type your company's name to make it yours. Free, and we write back within a week.

Write to us ↓

Research peer-reviewed at ACM MobileHCI 2026Open interface (HSI, Apache 2.0)Read by the founders, every time

Humans stay the boss.

AI does more of the work every month. It writes the code, drafts the reports, runs the workflows. It should still answer to the people it works for.

That takes two things most AI doesn't have: what your team already decided, and how the person in front of it is doing right now. That's what we build.

Synapse

Your team's memory and guardrails, for every AI agent you use.

synapse.work ↗

Life

Knows how you're doing from your own devices, and tells your AI how to pace itself. Shared only where you allow.

Get Life ↗
SDK

For builders

Our model and open interface, HSI, for companies building their own AI products.

Read the HSI spec ↗

One stack

A model, an interface, and products built on them.

We're a research company. The model learns how one person is doing. The interface carries that, with permission. Our products, and yours, put it to work.

ProductsBuilt on human state

Tools that change how AI works with each person.

HSI · Human State InterfaceCarries it, with permission

An open spec (Apache 2.0). One coarse reading, one way, only to what the person connects. Read the spec ↗

HSM · Human State ModelLearns each person, on their device

Signals from the phone, watch and computer people already use, read against their own baseline. Always an estimate, with its confidence.

What leaves the device

Your data stays. Only the state travels.

The Human State Model runs on the person's own phone, watch and computer. The only thing it can share is one coarse reading, sent through HSI to agents the person connects. Try to get anything else across.

On the person's device
  • Heart rate and HRVstays
  • Sleepstays
  • Typing rhythmstays
  • App switchingstays
Human State ModelLearns this person's baseline, here
What receives
HSI reading
ModeFocus Shield
Confidence0.8
Shared with
Raw signalsNone

Illustrative. The person picks who receives the reading and can switch it off at any time. Every reading is an estimate with a confidence level.

How it works

You write to us

Five minutes. How your company uses AI or technology today, and where it falls short.

We read it ourselves

Every note goes to the founders. We look at your work and where human context would help.

We write back

A short study and a demo built around your work, at your own synheart.to address, within a week.

Write to us.

A few sentences is plenty. The more specific, the better your study.

Israel GoytomReads every note
Rick TranReads every note
What brings you here? (pick any)
How many people would this touch? (optional)
We use this only to prepare your study and reply to you. No mailing list.

If you have more time

The science behind it.

Synheart is built on findings that have been studied for decades. Here are the ones we lean on most, with sources, so you can check them yourself.

  1. People hold about four things in mind at once.

    Working memory keeps roughly three to five meaningful chunks at a time. An AI can put fifty on screen.

    For your AI: when someone is stretched, one step at a time gets through. A wall of text gets skimmed.

    Cowan, “The magical number 4 in short-term memory”, Behavioral and Brain Sciences, 2001.

  2. Interrupted people work faster, and pay for it in stress.

    In a controlled study, people who were interrupted finished their tasks sooner but reported more stress, frustration, time pressure and effort.

    For your AI: a ping at the wrong moment costs more than the seconds it takes to read.

    Mark, Gudith & Klocke, “The cost of interrupted work: more speed and stress”, ACM CHI, 2008.

  3. What helps a beginner slows an expert down.

    Detailed, step-by-step guidance helps people who are new to a task and gets in the way of people who know it well. Researchers call it the expertise reversal effect.

    For your AI: the right amount of detail depends on the person and the moment, so it has to adapt.

    Kalyuga, Ayres, Chandler & Sweller, “The expertise reversal effect”, Educational Psychologist, 2003.

  4. Heart rhythm is tied to the brain’s control center.

    Heart rate variability is associated with prefrontal function, the part of the brain behind focus, planning and self-control.

    For your AI: a signal from the body can say something about how much room a person has right now, without reading a word they write.

    Thayer, Hansen, Saus-Rose & Johnsen, “Heart rate variability, prefrontal neural function, and cognitive performance”, Annals of Behavioral Medicine, 2009.

  5. Acute mental stress shows up in the heartbeat.

    A meta-analysis of short-term HRV studies found measurable shifts in heart rate variability when healthy adults were under acute mental stress.

    For your AI: a watch already on someone’s wrist carries a usable signal.

    Castaldo et al., “Acute mental stress assessment via short term HRV analysis in healthy adults”, Biomedical Signal Processing and Control, 2015.

  6. There is no universal normal.

    Heart rate variability differs widely from person to person, with age, fitness and more. A reading only means something against that person’s own baseline.

    For your AI: this is why Life learns each person’s baseline on their own device and never compares people.

    Shaffer & Ginsberg, “An overview of heart rate variability metrics and norms”, Frontiers in Public Health, 2017.

Our own work

  • Peer-reviewed
    Syni: Acting on Human State Without Sensing It. An on-device agent that adapts to a person’s state without ever seeing their raw signals. ACM MobileHCI 2026, late-breaking work.
  • Open spec
    Human State Interface (HSI) 1.3. The open contract our products speak: one coarse, consented, versioned reading of a person’s state that any system can check, with a public schema and test vectors under Apache 2.0. Whitepaper · GitHub
  • Preprint
    Pulse Focus. Validates the Focus Performance Score, a short attention test on the phone, on 111,133 trials from 466 people, with strong test-retest reliability. Its main component is linked to activity in the brain’s conflict-monitoring region on fMRI (N = 55). arXiv 2606.03164
  • Preprint
    Synheart Capacity. Estimates stress and effort from heart and skin signals in people the model never trained on: 70.0% and 72.2% balanced accuracy, against a 42.5% baseline (SWELL-KW). arXiv 2605.24416
  • Preprint
    Synheart Emotion. Detects stress from a wrist sensor alone, running entirely on the device: 74.5% accuracy (WESAD). arXiv 2511.06231

These studies support the direction, and none of them is a promise about any one person. Our readings are estimates that carry a confidence level, and they are never used to score anyone.