iQuve
Real-time video-based AI tutoring service for ages 4-6 - Production-level multimodal AI agent system
Overview
iQuve, an AI tutoring service for children aged 4-6 in which a character converses over live video and responds to what the child says and does. Ownership of the AI agent system and the backend that runs it.
Core Problem: A Four-Year-Old Neither Waits Nor Follows Instructions
This age group breaks the usual assumptions. Children speak unpredictably, go off-script constantly, and lose the thread when a response takes too long. A pipeline tuned for adult learners — wait for a full utterance, reason carefully, reply — simply fails here.
Three design directions followed: keep the interaction loop short enough to hold attention, pre-generate content to keep the live path thin, and design prompts that steer a wandering conversation back on topic without ever flagging the child as wrong.
Key Contributions
- Designed and implemented the core AI agent architecture — state, turn handling, and content flow behind the character
- Built the backend systems and real-time communication protocol — across client, agent, and media layers
- Selected models per task from LLM benchmarks — avoiding single-model dependency, placing high-capability models only where quality is perceptible
- Built the automated content generation pipeline — dynamic prompt scaffolding that produces lessons without hand-writing each script
- Developed the real-time video analysis and feedback loop — reacting to what the child is doing, not only what they say
Tech Stack
- AI: Gemini Pro, Gemini Flash, MiniMax API, Whisper API
- Backend: NestJS, TypeScript
- Infrastructure: Cloud GPU, Real-time Processing