Vision

One perception layer. Three ways it serves a life.

What Ethan AI is building, what it does today, and what it is for.

My son Ethan is non verbal and autistic. Like many families, we learned early that the hardest moments of his day arrive without warning, and that the only safeguard most families have is a pair of human eyes that cannot be everywhere.

I have spent seventeen years building video and vision systems. It took a long time to see that the problem in front of me at home was the same class of problem I had been solving professionally, and that nobody was solving it for families like ours.

The through line

Ethan AI builds a behavioural intelligence layer: a system that understands what a person is doing, continuously, from sensors already present in the room. Ordinary cameras. Ordinary audio. No wearables, no new hardware, nothing for the person to tolerate.

That single capability serves three different consumers, and those three consumers define our horizons. What changes across them is not the technology. It is who receives the output, and what they do with it.

One perception layer serving three horizons Ordinary cameras and ordinary audio feed a single behavioural intelligence layer. That layer produces three outputs: a log for a human analyst, delivering understanding, live today; an alert for a human responder, delivering intervention, in deployment; and an action for an actuator, delivering independence, a long horizon research direction. Ordinary cameras RTSP CCTV, IP, webcam Ordinary audio already in the room Behavioural intelligence layer Understands what a person is doing, continuously. No wearables. No new hardware. No child data used to train models. HORIZON 1 Log to a human analyst Continuous, time stamped behaviour records for ABC analysis. Understanding LIVE TODAY HORIZON 2 Alert to a human responder Safety critical recognition, where seconds decide outcomes. Intervention IN DEPLOYMENT HORIZON 3 Act to an actuator The perception layer assistive robotics is missing. Independence RESEARCH DIRECTION
One capability, three consumers. What changes across the horizons is not the technology. It is who receives the output, and what they do with it.

Horizon 1 · Log

Horizon 1 Live today

Log

The output goes to a human analyst.

Continuous, time stamped behaviour records that make ABC analysis possible on complete data rather than fragments reconstructed from memory. Therapists see what preceded a behaviour, not what someone recalled afterwards. Parents see that a pattern clusters around a particular transition, hour, or setting.

This is the unglamorous foundation, and it is the layer everything else is built on. Manual behavioural documentation is known to be incomplete, inconsistent between observers, and delayed. Automated capture addresses a problem the field has documented for decades and largely accepted as unsolvable.

This horizon is where our published evidence sits. An eight week study with 34 non verbal children, reported at NIMHANS in July 2026, measured 88% sensitivity and 84% precision against expert human observation, with caregiver burden falling 16% on the Zarit scale. Four behaviours run in production today, with a fifth in extended pilot. The full results are here, and the argument for why the antecedent is the weakest part of a manual record is set out here.

Horizon 2 · Alert

Horizon 2 In deployment

Alert

The output goes to a human responder, urgently.

Safety critical recognition, where seconds decide outcomes. A gas knob turned. A child alone in the kitchen. A hand reaching for a knife. A harmful substance being ingested. Leaving a safe zone.

The log answers what has been happening. The alert answers what is happening right now, and someone needs to come. These are the same perception capability under two different time constraints.

Of that list, safe zone exit is the one running in extended pilot today. The rest describe the class of problem this horizon covers, not a shipped feature list. We name them: they are what families actually ask us for, and being specific about the target is more useful than being vague about the roadmap.

Horizon 3 · Act

Horizon 3 Long horizon direction

Act

The output goes to an actuator.

Behaviour understanding is the missing perception layer for assistive robotics. A system that knows what a person is doing, and what they are trying to do, is the prerequisite for a system that can help them do it, completing a task without an adult being summoned.

This is a research direction spanning years, not a product roadmap. We name it: it is why the foundational work matters, and it answers the question every parent of a non verbal child carries and rarely says aloud: what happens when I am no longer here?

Children with special needs today are aging adults tomorrow. Care systems are not adapting fast enough. Independence, not supervision, is the only durable answer.

A behaviour recognised is a behaviour that can be logged, alerted on, or acted upon. The engineering investment, meaning perception accuracy, behaviour composition, and running continuously and affordably on existing hardware, compounds across all three horizons rather than splitting between them.

We build the layer. The horizons are what it becomes useful for, in sequence, as the accuracy and the trust are earned.

Work on this with us

Horizon 1 needs clinical researchers to co design validation: inter rater reliability, ecological validity, longitudinal stability. Horizon 2 needs therapy centres and families willing to define what a safety critical alert should mean in their setting. Horizon 3 needs people who think about assistive robotics.

hello@ethanai.in  ·  Shilpa Pakki on LinkedIn  ·  What runs today