AnakUnggul - ASD Caregiver Agent

Real-time AI copilot for autism caregivers: listens, observes, delivers calm guidance.

Comment

Inspiration

More than 1 in 36 children are diagnosed with autism spectrum disorder (ASD), yet the caregivers behind them — parents, grandparents, teachers — face crises largely alone. A meltdown can escalate in seconds. Finding the right clinic can take weeks. Writing a handover note for a therapist's appointment means recalling details under stress.

We built AnakUnggul because caregivers deserve a calm, knowledgeable companion at 2am when no therapist is available — one that speaks Indonesian and English, understands the real language of caregiving, and connects to the right local resources instantly.

What it does

AnakUnggul is an A2A-compliant ASD caregiver support agent built on the Prompt Opinion platform. It provides seven specialized skills in a single conversational interface:

  1. find_asd_resources — curated clinics, therapists, schools, and hospitals in Jakarta; live Google Search results for any other city worldwide
  2. suggest_interventions — evidence-based intervention strategies for a specific behavioral trigger
  3. get_de_escalation_steps — step-by-step protocol for active distress or meltdown situations
  4. assess_escalation_risk — risk level assessment from behavioral pattern descriptions
  5. draft_therapist_handover — structured clinical briefing note for therapy appointments
  6. assess_caregiver_wellbeing — caregiver stress screening and self-care guidance
  7. get_sensory_diet_plan — personalized sensory diet plan based on the child's profile

The agent automatically mirrors the caregiver's language (Bahasa Indonesia or English) and never asks clarifying questions before acting — a critical design requirement for high-stress caregiving moments.

How we built it

  • Agent framework: Google Agent Development Kit (ADK) with gemini-2.5-flash-lite
  • Serving: FastAPI on Google Cloud Run (asia-southeast1), Dockerfile-based deployment via Cloud Build
  • A2A compliance: Custom hybrid response builder to satisfy Prompt Opinion's parser — result.kind="task", result.task wrapper, status.state="TASK_STATE_COMPLETED" (ProtoJSON), and artifact parts with both kind and type fields
  • Agent card: Exposes both legacy PO fields (url, protocolVersion, preferredTransport) and supportedInterfaces, with explicit securitySchemes.apiKey
  • Security: API key middleware on all non-public endpoints; agent card public, / POST gated

Challenges we ran into

The biggest technical challenge was Prompt Opinion's A2A response schema. The platform requires a very specific hybrid payload that combines the legacy A2A spec with its own extensions — the exact required combination of fields took multiple iterations to discover (ProtoJSON enum strings vs. lowercase state, dual kind+type on artifact parts, the task wrapper shape).

A second challenge was ADK's tool schema behavior: the framework strips Python default parameter values from the generated JSON schema, causing the model to treat all parameters as required and ask clarifying questions before calling tools. We solved this by adding explicit defaults to all tool signatures and forceful per-skill directives in the system prompt.

Accomplishments that we're proud of

  • All 7 skills respond on the first turn without asking clarifying questions — even for vague, emotionally charged caregiver messages in Indonesian
  • Fully published and callable on the Prompt Opinion platform, passing the A2A smoke test suite
  • Bilingual (Indonesian/English) with automatic language detection
  • Real curated ASD resource database for Jakarta with live Google Search fallback for any city worldwide

What we learned

Building for a real caregiver population forced us to prioritize speed and empathy over feature completeness. A caregiver in a meltdown situation cannot wait for an agent to ask three follow-up questions. The system prompt and tool default design directly shaped the user experience more than the model choice did.

We also learned the practical realities of A2A platform integration: open standards still have implementation-specific quirks, and building compatibility requires reverse-engineering the consuming platform's parser expectations.

What's next for AnakUnggul - ASD Caregiver Agent

  • FHIR integration: Attach child health records context via SHARP extension so interventions are grounded in the child's documented history
  • Session continuity: Multi-turn memory so the agent can track behavioral patterns across weeks, not just respond to individual queries
  • Community insights loop: Surface anonymized patterns from the caregiver community to improve intervention recommendations over time
  • Expansion to more cities: Grow the curated ASD resource database beyond Jakarta to cover Surabaya, Bandung, Medan, and other Indonesian cities

Built With

Share this project:

Updates