AIdeas Finalist: Social Ace - Unlocking Connection for Autistic Youth through AI on AWS
Social Ace is an AI-driven social skills trainer transforming how autistic youth practice communication. As a Top 50 Finalist in the AWS 10,000 AIdeas Challenge, this project combines gamified dialogue with a high-performance AWS architecture—including Bedrock AgentCore and Nova Sonic 2. Our goal is simple: to provide an affordable, engaging, and secure "Social Gymnasium" where every child can build the confidence to thrive in the real world.
1: App Category
Social Impact
Social Ace is a dedicated Social Impact platform. We use inclusive innovation to help autistic youth practice essential skills, building the confidence they need for real-world connection.
2: Our Vision
Empowering the Silent Voice
For children with autism or language development delays, a simple conversation with a peer can feel like an impossible mountain to climb. They often struggle to initiate talk or maintain a topic, leading to social isolation.
Our vision is to transform this struggle into a journey of mastery. Social Ace is an engaging, AI-powered "Social Gymnasium" where youth can practice guided dialogue in a safe, judgment-free environment. By simulating daily life scenarios, we empower them to build the verbal confidence needed to navigate the real world.
3: Why This matters?
The Challenge: Clinical Barriers
Traditional social training (ABA and Speech Therapy) is effective but faces three critical hurdles:
- High Costs: Professional fees create a significant financial barrier.
- Low Accessibility: Shortages of experts lead to exhausting waitlists.
- Low Engagement: Repetitive, clinical methods can feel like a chore for children.
Social Ace bridges the gap between therapy and daily life with a 24/7 AI-driven "Social Gymnasium" that makes communication practice accessible, gamified, and rewarding.
Market Opportunity : Gap in Current Solutions
Most existing social tools for neurodivergent youth fall into two extremes, neither of which fully meets the need:
- Static Apps: Most digital tools use pre-recorded videos or multiple-choice buttons. These are linear and repetitive, failing to prepare children for the unpredictable nature of real-world conversation.
- Human-Led Therapy: While highly effective, traditional SST is resource-heavy. High costs and specialist shortages create massive barriers for rural and lower-income families.
Social Ace occupies the 'Sweet Spot' by combining clinical therapy with the accessibility of a web app.
Feature | Static Apps | Clinical Therapy | Social Ace |
|---|---|---|---|
Interaction | Pre-scripted / Buttons | Dynamic / Human-led | Dynamic / AI-led |
Availability | 24/7 | Restricted / Scheduled | 24/7 |
Cost | Low | Very High | Low / Scalable |
Feedback | Basic (Right/Wrong) | Detailed & Clinical | AI-Driven & Real-time |
4: Core Features and Therapeutic Framework
a. The Experience: Gamified Mastery
- AI Dialogue: Immersive, persona-based social scenarios.
- Smart Scaffolding: Contextual "Cue Cards" that provide hints before frustration begins.
- Real-Time Feedback: Gentle, constructive prompts for on-the-spot social correction.
- Mastery Journey: A gamified Token Economy (points/badges) to reinforce social wins.
b. The Science: Clinical Alignment
Social Ace is built on a rigorous ABA-inspired backbone to ensure measurable growth:
- Token Economy: A gamified rewards system where users earn points and badges for positive social behaviors, reinforcing real-world habits through digital incentives.
- Errorless Learning: AI-driven prompts minimize failure, building the consistent confidence needed for social growth.
- Social Reciprocity: The Evaluator Agent monitors the "back-and-forth" flow to ensure genuine interaction.
- Joint Attention: Tracks the user’s ability to stay focused on a shared, partner-led topic.
- Pragmatic Language: Gently corrects "topic hijacking" and impolite phrasing using standard Social Skills Training (SST) rubrics.
5: How I Built It
a. Application Components Overview
Frontend Website
Social Ace provides an interactive web interface built with Next.js and ReactJs. Users can register through AWS Cognito authentication, browse different social topics (e.g. music, sports, travel), and engage in real-time voice conversations with AI avatars. The frontend features a responsive design with Tailwind CSS and real-time WebSocket communication for bidirectional audio streaming.
Backend AI Agent
The backend consists of AI agents developed using the Strands Agent framework and deployed on AWS Bedrock AgentCore. The system employs multi-agent orchestration with two specialized agents:
- Conversationalist Agent: Handles real-time voice conversations using Nova 2 Sonic model
- Evaluator Agent: Analyzes social behaviors using DeepSeek R1 or Minimax m2.5 model for real-time feedback
The agent architecture supports session persistence, dynamic prompt retrieval, and gamification features including points tracking and badge progression (Bronze, Silver, Gold).
b. Application Architecture
Social Ace leverages a serverless architecture on AWS for scalability and cost-effectiveness:
Frontend Hosting
- AWS Amplify Hosting provides Git-based continuous deployment for the Next.js application
- Automatic builds triggered on every push to the main branch
- Built-in CloudFront CDN for global content delivery with automatic HTTPS
Backend Runtime
- AWS Bedrock AgentCore provides serverless runtime for the conversational agent
- Lightweight and cost-effective compared to ECS/Fargate
- Strong agentic capabilities with built-in WebSocket support
Data Layer
- DynamoDB stores prompts, user progress, and session history using a single-table design pattern
- Topic-specific conversation history isolation (each avatar maintains independent context)
- Automatic data expiration with 90-day TTL for privacy compliance
Serverless Functions
- AWS Lambda functions handle:
- Presigned URL generation for secure WebSocket connections (SigV4 authentication)
- Prompt retrieval from DynamoDB based on topic and age group
- User points and badge updates for gamification
- REST API proxy for hint and vocabulary features
API Gateway
- REST API with Cognito User Pool authorizer for JWT validation
- CORS-enabled endpoints for frontend integration
- Routes for WebSocket Presigned URL generation and chat assistant features
Authentication
- AWS Cognito provides OAuth2 authentication for the web frontend
- JWT tokens validated at API Gateway layer
- Presigned URLs with SigV4 signatures for AgentCore WebSocket connections
c. Agentic AI – Orchestration and Dialogue Evaluation
Multi-Agent Architecture :
Social Ace employs a sophisticated multi-agent system with specialized roles:
1. Conversationalist Agent (Primary Agent)
- Model: Amazon Nova 2 Sonic
- Purpose: Real-time voice conversation and orchestration
- Features:
- Bidirectional audio streaming via WebSocket
- Dynamic prompt retrieval based on topic and age group
- Session management with conversation history tracking
- Tool calling for hint and vocabulary generation
2. Evaluator Agent (Sub-Agent)
- Model: DeepSeek R1 (us.deepseek.r1-v1:0) or MiniMax 2.5 (minimax.minimax-m2.5)
- Purpose: Real-time social behavior analysis
- Features:
- Analyzes user responses against social behavior rubric
- Provides structured evaluation with confidence scores
- Triggers gamification updates via Lambda
3. Chat Assistant (Utility Agent - tools)
- Model: Amazon Nova Micro
- Purpose: Lightweight hint and vocabulary generation
- Features:
- Fast response time (<500ms)
- Cost-effective for frequent requests
- Generates 4-5 conversation hints or 10-15 topic-specific vocabulary words
Social Behavior Evaluation Logic :
The Evaluator Agent uses DeepSeek R1 to analyze user responses against a comprehensive rubric:
- Positive Behaviors (+1 point each, identify at most ONE):
- Joint_Attention: Acknowledging and building on the conversation partner's topic
- Social_Initiation: Starting a new relevant topic or asking engaging questions
- Conversational_Reciprocity: Balanced back-and-forth exchange with appropriate turn-taking
- Inquiry_Skills: Using Wh-questions (who, what, where, when, why, how) to show interest
- Topic_Maintenance: Staying on topic for multiple turns and developing the conversation
- Active_Listening: Using bridge phrases ("That's interesting!", "Tell me more") to show engagement
- Corrective Behaviors (-1 point each, identify at most ONE):
- Non_Contextual_Talk: Introducing topics that are clearly and completely unrelated with no connection to the current conversation
- Maladaptive_Language: Using rude, mean, or socially inappropriate language
- Repetitive_Speech: Repeating the exact same phrase or question 3+ times without variation
- Topic_Hijacking: Forcefully changing the subject without acknowledging the current topic
- Information_Monologuing: Long speeches without allowing turn-taking or checking for interest
- Interrogative_Overload: Asking multiple questions in rapid succession without waiting for answers
- Evaluation Flow:
- User completes a conversational turn
- Evaluator Agent analyzes the response with conversation context (last 5 turns)
- Identifies at most ONE behavior (positive, corrective, or none)
- Returns structured evaluation with confidence score
- Points Lambda updates user progress in DynamoDB
- Frontend receives real-time feedback via WebSocket
Agent Orchestration Framework:
The system uses the Strands Agent framework for agent orchestration:
Why Strands?
- Lightweight and simple architecture
- Excellent for rapid development and prototyping
- Built-in support for bidirectional voice streaming (i.e. BidiAgent)
- Easy integration with AWS Bedrock models
- Minimal boilerplate code
Model Selection Rationale:
Model | Purpose | Key Characteristics |
|---|---|---|
Nova 2 Sonic | Primary conversation (Speech to speech model) | Optimized for conversational AI with low latency and natural voice synthesis |
DeepSeek R1 or Minimax 2.5 | Social behavior evaluation | Excellent reasoning capabilities for nuanced analysis (~1-2s response time) |
Nova Micro | Hint/vocabulary generation | Ultra-fast and cost-effective for simple text generation (<500ms) |
This multi-model approach balances performance, cost, and quality across different use cases within the application.
This multi-model approach balances performance, cost, and quality across different use cases within the application.
d. Security Design: Safety & Child Privacy
Social Ace uses Amazon Bedrock Guardrails to create a "fail-safe" environment, prioritizing child safety and natural conversation.
- Jailbreak Protection (BLOCK): Uses a medium-sensitivity filter to block prompt injection and instruction overrides. Safety is non-negotiable.
- Content Moderation (DETECT): Monitors for inappropriate topics (violence, hate speech). To avoid distressing "hard stops," the Evaluator Agent handles corrections via educational feedback and point deductions.
- PII Privacy (MASK): Automatically redacts names, addresses, and phone numbers. Sensitive data is masked before storage, ensuring PII never reaches DynamoDB or CloudWatch logs.
- Fail-Safe Continuity: Implements a 3-second timeout and "fail-open" logic. Infrastructure glitches never interrupt a child's learning flow.
e. Key Technical Decisions
1. Serverless Architecture
- Why: Cost-effective, auto-scaling, minimal operational overhead
2. Multi-Model Strategy
- Why: Optimize for specific use cases (conversation vs. evaluation vs. utility)
3. Single-Table DynamoDB Design
- Why: Efficient queries, reduced costs, simplified data model
4. Presigned URLs with SigV4
- Why: Secure WebSocket connections without exposing credentials to frontend
5. Topic-Specific History Isolation
- Why: Different avatars maintain independent conversation contexts
e. Development Process – Spec-Driven Development with Kiro
My Development Approach:
I make use of Kiro and employed a Spec-Driven Development process, breaking the project into six granular, verifiable milestones to ensure technical excellence and testability:
- Phase 1: Foundation. Developed a workable voice-chat agent using Amazon Nova Sonic 2 as the core engine.
- Phase 2: Intelligence. Implemented a multi-agent orchestration between a Conversationalist Agent and an Evaluator Agent.
- Phase 3: Persistence. Built a long-term memory strategy for user recognition and conversation continuity across sessions.
- Phase 4: Interface. Developed the frontend web application using Next.js and React.
- Phase 5: Secure Asset Delivery. Integrated S3 Presigned URLs to ensure safe, temporary access to media and session data.
- Phase 6: Safety & Privacy Refinement. Deployed Amazon Bedrock Guardrails and PII masking to safeguard child privacy and enforce therapeutic boundaries.
Kiro helps me create a comprehensive set of requirements, design specifications, and technical details for each phase. The benefit of this approach is to ensure that every milestone is granular and verifiable, making it easier to monitor progress, conduct rigorous testing, and guarantee the technical excellence of the Social Ace platform.
5: Demo
Screenshot:
6: What I have learnt?
The Finalist Pivot: Putting Safety First
Reaching the Finalist stage brought a vital reminder from the judges: when building for children, privacy and security are everything. Following COPPA (child privacy laws) became my top priority.
This feedback shifted my focus from just "making it work" to making it safe. I integrated Amazon Bedrock Guardrails to transform Social Ace into a secure vault for our users. This experience taught me that for AI projects with social impact, responsibility isn't an extra feature—it’s the foundation.
Key Insights on AI-Driven Development
- Spec-Driven Development (SDD) vs. "Vibe Coding" Ad-hoc "vibe coding" leads to technical debt. I learned that SDD is the only path to professional, scalable software. Clear documentation and granular milestones are what turn a "cool demo" into a sustainable product.
- The Human-AI "Handshake" Partnering with Kiro (my AI dev partner) taught me that the Specification Phase is the most critical. The "handshake"—the iterative process of questioning and clarifying AI-generated specs—is where the real engineering happens.
- Maintaining a Critical Eye AI can suggest technically incompatible designs. Always validate AI suggestions with independent research before locking in your architecture.
- Driving Early "Testable" Deliverables I prioritized UI-ready features from Day 1. Visualizing the MVP developed by AI early allowed me to catch design flaws and user experience hurdles before they became expensive backend problems.
The Evolution of the Software Engineer
This journey confirmed that AI is not strictly replacing engineers—it is elevating them. My role has shifted from "Coder" to "Lead Architect." In this new era, your value lies in your ability to maintain the high-level view, think critically, and act as the ultimate authority on quality. AI provides the bricks; the human provides the blueprint.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article