Back to projects

Co-Regulation Hub

A multi-agent system that notices when a sensory-sensitive child is becoming overwhelmed and, in real time, coaches their caregiver through a clinically-approved calm-down plan. Built for the Endava Agentthon on Google ADK and Gemini.

Google ADK Gemini 2.5 Flash Vertex AI Multi-agent

The problem

Caregivers of children with autism and sensory sensitivity often can't tell early enough that a child is escalating — and under stress, they struggle to recall the specific steps their therapist prescribed for that child. The Hub surfaces the right plan, at the right moment, in plain language, so the caregiver can act with confidence and consistency.

How it works

Sense → protect → understand → coach

A root orchestrator delegates each alert through a fixed pipeline of specialist agents. Every stage hands clean, verified context to the next, and nothing reaches the caregiver until it's grounded in the child's own care plan.

1
Sensebio_sentryacoustic_core
Two sensor agents establish a baseline and flag an event when biometrics or ambient noise rise 30% above it.
2
Protectdata_defender
Strips PII, redacts voice audio, and validates Australian privacy compliance before anything moves on.
3
Understandclinical_context_agent
Ingests the child's support plan and cross-references the schedule to confirm a genuine event, not a planned high-activity lesson.
4
Coach & logpixel_buds_vox
Delivers a sub-25-word whisper cue to the caregiver's earbuds — or a 5-step plain-language plan if none exists — then records the outcome for clinicians.

A walk through one alert

How a single event flows through the pipeline

An illustrative trace of one escalation, from raw sensor anomaly to the cue that reaches the caregiver. In the MVP the sensors and hardware are simulated so the full flow can be demonstrated end to end.

ORCHESTRATOR Alert received. Running sensor sweep…
BIO_SENTRY HR 112 bpm vs 80 baseline (+40%) — flagged ANOMALY
ACOUSTIC_CORE Ambient 82 dB vs 45 baseline (+82%) — distress frequencies detected
DATA_DEFENDER PII stripped, voice redacted. Region pinned: australia-southeast1
CONTEXT Schedule = "Reading" — not a high-activity lesson. Confirmed genuine event
CLINICAL PBSP matched: auditory trigger. Protocol → headphones / weighted lap pad / quiet space
PIXEL_BUDS_VOX Cue → "Sarah's overwhelmed by noise. Offer her headphones and the quiet corner, and keep your voice low." (Buds: Transparency)
LOG Incident summary written for therapist review
5
Specialist agents
30%
Anomaly threshold over baseline
<25
Words per whisper cue
0
PII fields logged

Under the hood

Tools & technologies

A Google-Cloud-native agent stack, built around clinical safety and Australian data-residency requirements rather than a generic chatbot.

Agent framework

Google ADKMulti-agent orchestrationSub-agent delegationTool-calling

AI / LLM

Gemini 2.5 FlashVertex AI

Google Cloud

Vertex AIBigQueryaustralia-southeast1

Knowledge / RAG

PBSP dossier ingestionWorkspace file scanConditional protocols

Integrations

Google Calendar (MCP)Pixel Buds Pro 2Wearable biometrics

Privacy & compliance

PII redactionVoice-audio redactionAU data residency

Tooling

Pythongcloud CLIGemini CLIadk web

Project highlights

What the system gets right

Real-time escalation detection
Reads biometric and acoustic signals and flags an event when either rises 30% above the child's own baseline.
Context-aware false-alarm filtering
Cross-checks the classroom schedule so a spike during music or sport isn't mistaken for distress.
Privacy by design
Strips PII and redacts voice audio before anything is logged, with data pinned to the Australian region.
Grounded in the child's own care plan
Retrieves the therapist-authored support plan — triggers, diagnosis, medications, approved protocols — and acts only on what's been clinically approved.
Calm, concise caregiver coaching
Delivers a whisper cue under 25 words to the caregiver's earbuds in Transparency mode, so they stay connected to the child.
Closes the loop with clinicians
Writes a structured incident summary to the allied-health data store for therapist and NDIS review.
Safety guardrails
Stops instantly on "Deactivate", "Stop Monitor", or "Privacy Mode", directs to 000 for emergencies, and responds in the caregiver's language.