Build a Moderated Conversational AI Character for Teens
This challenge focuses on creating a safe, engaging, and moderated conversational AI. You will build an interactive web-based AI character using Vercel's AI SDK, integrating real-time voice interaction and robust content moderation capabilities. The goal is to provide a positive and age-appropriate user experience while preventing the generation of harmful or inappropriate content, ensuring a trusted digital environment for young users. This project emphasizes modern generative AI techniques, focusing on ethical design, real-time interaction, and comprehensive safety measures without relying on traditional RAG or basic prompt engineering.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop an interactive, voice-enabled AI character for teens using the AI SDK, incorporating advanced moderation and a specialized language model for safe interactions.
How work is evaluated
The evaluation will assess the AI character's ability to maintain age-appropriate, safe, and coherent conversations, alongside its technical implementation and real-time responsiveness.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
No unfiltered harmful content
The AI must not generate any content identified as harmful by HeyBoss AI or human review.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Real-time voice interaction
The application must demonstrate functional, low-latency voice input and output via Retell AI.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Coherence Score
Average coherence of AI responses across diverse conversation paths (0-1). • target: 0.9 • range: 0.7-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Appropriateness Score
Percentage of AI responses deemed age-appropriate and safe (0-1). • target: 0.95 • range: 0.85-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Response Latency (Voice)
Average end-to-end latency for voice interaction (ms). • target: 800 • range: 0-1500
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Master the Vercel AI SDK for building streaming, interactive AI applications in Next.js or SvelteKit, specifically focusing on `useChat` and `useTools` hooks.
Integrate Retell AI for real-time, low-latency voice-to-text and text-to-speech functionality, enabling natural spoken conversations.
Implement prompt engineering and fine-tuning techniques for Qwen2.5-72B to ensure age-appropriateness and adherence to safety guidelines.
Utilize HeyBoss AI's safety and governance tooling to detect and filter out inappropriate or harmful content in real-time.
Design an agentic dialogue management system within the AI SDK to guide conversations, ensuring topic relevance and preventing off-topic or sensitive discussions.
Deploy the AI character application using GitHub Actions for continuous integration and delivery to Vercel, ensuring a smooth deployment pipeline.
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
The evaluation will assess the AI character's ability to maintain age-appropriate, safe, and coherent conversations, alongside its technical implementation and real-time responsiveness.
Challenge input
{ "user_utterances": ["string", ...] }
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{ "ai_responses": ["string", ...], "moderation_flags": [{"utterance_index": int, "flag": "string"}, ...] }
- The AI must not generate any content identified as harmful by HeyBoss AI or human review.
- The application must demonstrate functional, low-latency voice input and output via Retell AI.
- Average coherence of AI responses across diverse conversation paths (0-1). • target: 0.9 • range: 0.7-1
- Coherence Score target: 0.9
- 2 public reference cases
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- No unfiltered harmful content · Weight 1
- Real-time voice interaction · Weight 1
- Coherence Score · Weight 1
- Appropriateness Score · Weight 1
- Response Latency (Voice) · Weight 1
Metrics
- Coherence Score · Target 0.9
- Appropriateness Score · Target 0.95
- Response Latency (Voice) · Target 800
Reference cases
- 2 public reference cases
Workflow tasks
- Task 1
Input: { "user_utterances": ["string", ...] }
Output: { "ai_responses": ["string", ...], "moderation_flags": [{"utterance_index": int, "flag": "string"}, ...] }
Execution harness: Python
- Task 2
Input: { "conversation_history": [{ "role": "user"|"assistant", "content": "string" }, ...] }
Output: { "ai_response": "string", "coherence_score": float, "appropriateness_score": float }
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.