Building a real-time AI tutor for 5-year-olds

6 hours ago

Sub-second responses are essential for keeping children engaged in learning; long pauses cause attention to wander.
The standard AI agent tool loop was abandoned due to latency issues, with custom harness enabling streaming of multiple actions for real-time interaction.
Two agents were implemented: a 'converser' for immediate interaction and a 'planner' for pedagogical reasoning, running asynchronously to anticipate child's actions.
Safety checks run in parallel with generation to avoid delays, using an LLM classifier to block inappropriate content without slowing responses.
Anticipating child responses allows pre-generation of answers to closed-ended questions, reducing wait times.
The architecture emphasizes balancing instruction following, latency, and action flexibility, requiring custom observability and tracing.
Key lessons include owning the interaction loop, leveraging child's thinking time for planning, and gating safety on execution rather than generation.

Hasty Briefsbeta