Ethan AI: Agentic behaviour intelligence and alerting for autism care.

Introduction

Continuous behavioural observation supports individuals on the autism spectrum and their care teams across home, school, and therapy settings. In India, access to such observation is limited by the cost of trained personnel, and the caregivers who do this work are often on their feet for long hours and carry significant fatigue. The behaviours most relevant to daily participation and wellbeing — including self-injury, distress responses, stimming, and postural patterns — are precisely those that benefit from sustained observation across settings, where human attention is hardest to keep continuous.

Method

Ethan AI is a computer vision and audio model that runs on video sources already present in care environments — RTSP-enabled CCTV, IP cameras, and standard webcams — without requiring specialised hardware or autism-specific model training.

The architecture composes pretrained perception models:

• MediaPipe Pose for keypoints
• YAMNet for audio events
• SigLIP for image-text gating
• A vision-language model on an asynchronous channel for summarisation

A three-tier runtime separates continuous pose and audio processing (T0) from sampled image-text checks (T1) and asynchronous structured reporting (T2). The system’s central design element is a Behaviour Catalog in which each behaviour is specified as a composition of signals from the adapter set, so new behaviours are added as recipes rather than as separately trained models.

Results

The system detects head banging, glasses presence, distress vocalisation, and head posture. Preliminary internal evaluation indicates approximately 80% detection accuracy on head posture and 90% accuracy on glasses presence. Poor accuracy results and constraints on five additional behaviours — including hand flapping, rocking, falling, hair pulling, and climbing furniture — are discussed in the full abstract.

Discussion and Conclusion

Ethan AI is positioned as a behavioural observation aid that supports caregivers and clinicians — not a replacement for either.

The authors invite clinical researchers in global autism care to co-design validation studies on Ethan AI model reliability against trained observers, and on the stability of detection accuracy for Catalog-defined behaviours over time. Access and code details accompany the submission.

Keywords

Meet us at NIMHANS

If you are a clinician, researcher, parent, therapy centre, or school interested in behavioural observation for autism care — we would like to talk to you.

Ethan AI is a computer-vision and audio behavioural observation platform for autism care. See our product and how it works pages for the full picture, and our autism conferences directory for other events across India, Asia, and globally.