Inspiration
In Indonesia, there are fewer than 200 certified ASD therapists for a population of 270 million. For a caregiver in Makassar or Medan, the waiting list for a single consultation can stretch to months. Meanwhile, meltdowns don't wait.
The idea for AnakUnggul came from a simple, uncomfortable question: what happens at 2am when a child is in crisis and no professional is reachable? Most caregivers face this alone — overwhelmed, undertrained, and isolated. We built AnakUnggul to be the calm, knowledgeable presence that shows up when no one else can.
What it does
AnakUnggul is a real-time AI companion for caregivers of autistic children. It provides seven specialized skills through a single conversational interface:
- Intervention strategies — evidence-based responses to specific behavioral triggers (sensory overload, demand avoidance, shutdown)
- De-escalation protocols — step-by-step guidance for active meltdown situations
- Escalation risk assessment — evaluates severity from behavioral descriptions and flags when professional help is needed
- Therapist handover notes — structured clinical briefing documents for therapy appointments, generated from caregiver descriptions
- Caregiver wellbeing screening — stress assessment and personalized self-care guidance for caregivers under chronic pressure
- Sensory diet planning — personalized sensory activity schedules based on the child's sensory profile
- Local ASD resource discovery — curated database for Jakarta, live search results for any city worldwide
The agent automatically detects whether the caregiver is writing in Indonesian or English and responds in kind — no setup, no language selection, no friction.
It also runs a live multimodal session mode: caregivers can share real-time audio and video during an active crisis, and the AI observes behavioral patterns and delivers guidance while the situation unfolds.
How we built it
The core stack:
- AI: Google Gemini Live (real-time multimodal) + Gemini 2.5 Flash Lite (text reasoning), Google ADK for agent orchestration
- Behavioral models: Two custom Keras/TensorFlow models — an audio pattern extractor and a behavior extractor — trained on ASD behavioral data and deployed on Cloud Run
- Backend: FastAPI on Google Cloud Run (Asia Southeast 1), Firestore for session persistence and memory context
- Mobile: Flutter app with real-time WebSocket connection to the AI backend
- Agent layer: A2A-compliant agent published on the Prompt Opinion marketplace, exposing all 7 skills to other AI agents and platforms
- Community intelligence: A Moltbook agent (anakunggul) that publishes ASD caregiving education to the AI agent community and harvests community insights back into the platform
Everything runs on a serverless, auto-scaling architecture — zero fixed infrastructure costs when idle, scales to meet demand.
Challenges we ran into
The 2am problem is also a model problem. Caregivers in crisis don't write clean, structured queries. They write "dia nangis terus dan mukul2 kepala" ("he keeps crying and hitting his head"). The AI had to learn to extract structured clinical context from emotionally charged, fragmented, bilingual input — and respond with calm, actionable guidance rather than asking clarifying questions.
Real-time behavioral observation was technically the hardest part. Streaming audio and video through a live WebSocket session, running inference on behavioral patterns, and injecting observations into the Gemini Live context — all without adding perceptible latency — required careful async orchestration and fallback handling.
Building for a population with no margin for error. Bad advice during a meltdown can escalate the situation. Every response goes through safety layers: the system never diagnoses, never prescribes, and always flags when professional intervention is needed.
Accomplishments that we're proud of
- 11 real caregiving sessions logged since beta launch with active families — not test users
- The agent responds in under 3 seconds on average, in the correct language, without ever asking "can you tell me more?"
- Fully A2A-compliant and live on the Prompt Opinion marketplace — callable by any compatible platform or agent
- A community presence (anakunggul) that has built genuine engagement in the AI agent ecosystem around ASD caregiving topics
- An architecture that costs near zero when idle but can scale to thousands of concurrent sessions
What we learned
The most important lesson: the interface is the intervention. When a caregiver is in crisis, every extra tap, every clarifying question, every second of latency is friction that erodes trust and usefulness. Designing for the worst-case emotional state of the user — not the average case — completely changed how we built the agent's behavior.
We also learned that building for an underrepresented population requires going to them. The behavioral patterns we trained on, the triggers we modeled, the local resources we curated — none of that came from public datasets. It came from listening.
What's next for AnakUnggul AI
- FHIR integration: Connect with clinic EHR systems so the agent's guidance is grounded in each child's documented health history
- Clinic and community partnerships: Formal collaboration with ASD therapy centers and parent communities — we provide anonymized behavioral insights and usage data; they provide ground-truth feedback that improves the models
- Expanding the population: The core architecture is condition-agnostic. Next: adults with late-diagnosed ASD (a severely underserved group) and elderly individuals with Parkinson's or dementia, where caregiver support follows the same pattern
- Research layer: Structured research partnerships with universities and pediatric hospitals to build the first large-scale Indonesian ASD behavioral dataset
- Offline mode: Lightweight on-device inference for caregivers in areas with unreliable internet — because the 2am problem doesn't only happen in cities with good connectivity
Log in or sign up for Devpost to join the conversation.