Watching over, never watched

A privacy-first AI presence — for therapy, and for safety, at home.

Adam's Space AI turns careful attention into structured therapy records and gentle safety signals. It stays close to the people who need it, without ever putting them on a screen for anyone else to see.

Every calculation happens on the device in the room. Each frame lives in memory for up to two seconds and is never stored or sent — only events leave the device, and only to the person already responsible for that care.

See how it works ↓ Partner with us →

≤ 2 s

A frame stays in memory, then it's erased

Events only

What leaves the device — never video

2 tracks

One shared AI core

The problem

Three quiet crises, one root cause

In our experience, a therapist can spend up to a third of their time on documentation instead of watching a child learn. More than half of U.S. counties have no certified behavior analyst at all — for millions of families, therapy simply isn't somewhere they can drive to. In many other countries the situation is far worse — and it's not only about physical access, but financial access too. And for the parents of a child who communicates little, or not at all, there's a specific kind of fear that never really goes away: knowing something is wrong, without knowing what, until it's too late to help.

We didn't invent this problem. We're building the tool we wished existed while living it.

up to 1/3

of a therapist's time goes to documentation — in our experience

Time that could have gone to watching a child learn instead goes to writing it down after the fact.

54%

of US counties have no BCBA (Yingling et al., Autism, 2021)

For millions of families, certified therapy isn't a scheduling problem — it's a place that doesn't exist nearby.

no channel

for a child who can't say what's wrong

Knowing something is wrong without knowing what, until it's too late to help, is a fear that never fully goes away. Nearly half of autistic children try to leave a safe place (49%, Anderson et al., Pediatrics, 2012), and autistic people face a ≈2.5× higher risk of early death (Hirvikoski et al., BJPsych, 2016).

Scale

Autism around the world

The WHO estimates autism at 1% of the world's population — and that number shouldn't be read literally, country by country.

of the population, per WHO estimates

61.8M

people worldwide

1 in 127

is the chance of meeting an autistic person

Map of diagnosed autism prevalence by country worldwide

Diagnosed autism prevalence by country — a map of where autism is recognized, not of where there's more of it

By WHO estimates, autism affects 1% of the world's population — 61.8 million people, one in 127. Where children are systematically screened, the diagnosed rate is far higher: 1 in 31 children in the US, or 3.2% (CDC, 2025). The difference isn't statistical noise — it's a difference in who gets noticed.

The autism prevalence map by country doesn't show where autism is more common — it shows where it's better recognized. The darkest countries on the map — Japan, Brunei, Singapore, South Korea — aren't more "autistic" than the lighter Bangladesh, Nepal, Bhutan, India. They simply have a higher capacity to detect it: broader diagnostic criteria, more specialists, more attention to "milder" points on the spectrum.

Much of the growth in recorded cases isn't more autism; it's more noticed autism. Even in France, a country with one of the strongest healthcare systems, only 15% of an estimated 700,000 autistic people have a formal diagnosis.

This is exactly why we're building task-shifted access, not just a product for people who already have a therapist. The light part of the map isn't a place without autism. It's a place where there's simply no one to say so yet.

The architecture

Two problems. One architecture.

Everything runs on the same foundation: on-device computer vision, nothing leaving the room, a person always in the loop. What changes is what it's for — documenting a therapy session, or noticing that a child is heading for a door they shouldn't open. We start with the cameras of ordinary phones and tablets: therapy runs on a regular tablet or phone, so a clinic buys no hardware to begin. For safety, and later for clinic rooms, we're building our own stationary camera-based device: it sees at night and raises the alarm even without internet.

Why now. In 2026, for the first time, four things line up: an ordinary phone or a low-cost chip can recognize behavior from video in real time; models can be trained without huge datasets; regulators and clients demand that video never leave the room; and care is facing an acute staffing crisis. Only a few years ago, recognizing behavior from video reliably still meant sending it to the cloud — and no one lets a cloud camera into a child's home or a care home.

Track 1 — Therapy Support

Documentation first.

On a regular tablet. A real path to families who have no other access to therapy at all, tomorrow.

Track 2 — Safety

A gentle presence.

For children at home, for adults in supported living, and for older people in care homes. Never a security camera.

Track 1

Therapy Support

MVP · January 2027

Documenting the session

It watches the trial, not the clock. Prompted responses, independent ones, what changed since last time — structured automatically, exported as PDF or CSV from day one, with direct export into CentralReach during 2027. Runs on a regular tablet. MVP in January 2027; first clinics in the US from March 2027.

Helping the therapist see the pattern

Weeks of sessions become one picture: what's working, where progress has stalled. A second set of eyes on data the therapist already owns — never a decision made for them.

Later

Reaching the families who have none

Where there is no certified therapist within reach, a parent or a trained paraprofessional runs the session with our structure, and a therapist reviews the results remotely — the same task-shifting model the WHO has used for decades to stretch scarce specialists further. This is not AI replacing a clinician. It's a way to put a clinician's judgment where it currently can't reach.

Track 2

Safety

Pilots from January 2027

Warning about the things that matter

A door or window opening at night. A fall, or lying on the floor for too long. A moment of self-injury, or sudden chaotic movement. The signal goes first to the people already responsible — a parent or family member — and, if no one answers, to on-duty staff, with the person's consent. Nothing about this is built to catch anyone doing something wrong — it's built so a family can feel present even when they can't be in the room. The alarm works even when the internet is down.

Next · from late 2027

Seeing the warning signs before the storm

Early signs of distress — rising tension in movement, posture and voice — and over weeks, something more: patterns connecting behavior to schedule, to sleep, to meals, and changes in behavior that may mean pain. For a child who can't say what's wrong, this is sometimes the only way anyone finds out. We're careful here on purpose: these are hypotheses for a parent to bring to a doctor, never a diagnosis, never medical advice from a small sample.

Pilots from January 2027

Supported living, done right

The same architecture, deployed with community-based supported-living providers — first pilots in Europe in 2027, then in the US. Homes, not institutions. It lets one caregiver support more people safely, without turning someone's home into a place that feels watched.

Privacy & ethics

Privacy isn't a feature we added. It's the reason the architecture looks the way it does.

Each frame is processed in the device's memory and erased within two seconds — never stored, never sent anywhere. What leaves the device is a number, a trend, an alert: never a frame. Our models are trained on recordings of adult actors who said yes, specifically, in writing — never on footage of children, or of the homes and clinics we serve. This is not the same category of technology as public-safety or child-protection surveillance systems. We build for one thing: a private relationship of care — a family, a therapist, a caregiving team — never detection on behalf of an authority.

A frame stays in the device's memory for up to two seconds, then it's erased.

Models learn from consenting adult actors — never from children's footage.

The system assists people — it never replaces human supervision, never diagnoses, and never reports a family.

Built for care, never for surveillance — of residents, families or staff.

We read posture, not faces — and will test across body types, ages and lighting before we claim accuracy.

Patterns are hypotheses to bring to a doctor — never medical advice.

Roadmap

Where we are, where we're going

Two tracks, moving in parallel, sharing one foundation. We give dates as targets, not promises — every stage we list ends in something you can actually see: a published report, a working pilot, a real export.

Track 1

Now

Documentation prototype

Jan 2027

MVP on a regular tablet

Mar 2027

First clinics in the US — export to PDF and CSV

2027

Direct export into CentralReach; analytics

Later

Pattern view; task-shifted access for underserved regions

Track 2

Now

Studio dataset with consenting adult actors

Jan 2027

Pilots in Europe on phones and tablets — families and supported-living providers

Q2–Q3 2027

Safety in the US on phones and tablets; our stationary device — prototype and pilot batch

Late 2027

Early signs of distress

2028

Certified device (CE, FCC)

Who it's for

Six kinds of people already waiting on this

Families

Warnings for dangerous behavior, and a clearer sense of what's happening for a child who can't quite tell you. Free during the pilot.

ABA therapists & centers

Less time writing, more time watching a child actually learn. Your existing tools, just with less busywork feeding them.

Communities with no local specialist

A parent or a trained helper runs the session; a therapist reviews it from wherever they are. Not a replacement — a way to reach the families who currently have no one.

Supported-living operators & care homes

One team, safely responsible for more people — at night too — without a single camera anyone has to feel watched by.

Researchers & policy partners

A dataset built on consent from day one, and a methodology we're willing to publish, including the parts that don't work yet.

Architects, developers & property managers

We're developing Care-in-Architecture, our own standard for inclusive housing: homes where older people, families with children and neurodivergent people live independently, close to each other, with on-device AI as a quiet safety net — a monthly service the property manager can offer.

Evidence & honesty

What we know, and what we're still finding out

This section will keep changing as we learn more, on purpose. We'd rather be exact about what's real than impressive about what isn't.

Working today

A working documentation prototype. MVP in January 2027.

An on-device architecture that's actually built, not just designed.

A training process built on consenting adult actors — never on children's or clients' footage.

Accepted into NVIDIA Inception; our first application for a compute grant has passed the initial screening — the grant itself is not yet awarded.

Not proven yet

Accuracy in real homes and clinics — that's what our 2027 pilots exist to test, and we'll publish the results either way.

Dangerous-behavior detection is not yet validated — homes and supported living will be tested separately: different setting, different population.

Patterns linking behavior to sleep and routine are future hypotheses, never medical advice.

Savings of 20–40% on therapy and supported-living costs are our model estimate, not yet measured. A reference point: Swedish municipalities replace routine night visits with digital night supervision, and their association SKR estimated national savings of about SEK 1.4 billion a year.

Team & origin

Why we're building this

This started with our own children. Two of them are autistic — one speaks several languages, one doesn't speak at all. How much of their therapy time went into writing things down instead of teaching, and how impossible it is to track the precursors of a meltdown or dangerous behavior by hand, 24/7, without turning our own presence beside the child into constant pressure on him — that's what started all of this.

Our children showed us what to build — not what to train on: the models themselves learn from consenting adult actors, never from footage of children. We went into this deliberately, with our eyes open — understanding in advance where this work could lead, not stumbling onto it along the way. Behind it is a specific goal: to make life better for autistic people in general, and to build a real, happy, thought-through model of supported living — for Nika and Adam, not for some abstract idea somewhere in the future. This isn't a side project or an adjacent hypothesis. It's the future we're building for our own children — and for many other families alongside them.

Oksana (Ksenia) Novitskaia

Founder & CEO · methodology

LinkedIn ↗

Egor Rebrikov

CTO

LinkedIn ↗

Andrey Pirogov

CMO

LinkedIn ↗

Alexandra Primaikina

ABA therapist

Vasily Lomakin

ML Lead

Partner with us

Three ways to work with us

Grants & funders

We're seeking pilot funding to validate both tracks on real, expanded, consented data, and to publish what we find either way.

Pilot sites

ABA centers, schools, community-based supported-living providers, care homes: run this alongside what you already do, at no cost during the pilot, with full data protection worked out upfront.

Investors

Early, mission-first, privacy-first, two tracks on one foundation, with a clear path to real B2B revenue in a $26.6 billion market for ABA services and AI monitoring in care — part of a long-term care market worth more than $1.2 trillion a year.