Building a real-time AI tutor for 5-year-olds
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- The development team set out to build an AI tutor for children ages 4-9 that teaches math and reading, making sub-second response a non-negotiable constraint because a 2-second pause is enough to lose a child's attention and stop learning.
- The custom harness streams multiple actions in a single model response and uses an interpreter to execute each action while the model continues generating, so the child waits only about 30 tokens for the first action instead of the full response.
- The converser-planner split separates the agent into a real-time "converser" that interacts with the child and an asynchronous "planner" that reviews the conversation against lesson objectives during the child's thinking or talking gaps, sharing state through an append-only event log so neither waits on the other.
- Closed-ended question handling pre-generates responses to likely answers on branched trajectories forked from the trajectory log, then matches the child's answer to a branch and plays the response without waiting on a fresh model call.
- Playtests revealed the latency stakes directly: one six-year-old girl figured out she only needed to pay attention part of the time and could still keep up — the team called this "the moment she stopped learning" — while a six-year-old boy simply asked when the agent was done thinking.
- The safety classifier, an LLM taking roughly 500-1000ms, runs in parallel with a small model generating an "eager response" that mirrors the child, gating execution rather than generation to avoid a latency hit on every turn; flagged turns discard the reflexive response and route to guidance that acknowledges the issue without echoing harmful content.
- Tradeoffs the team flagged include cost from the planner running on a more capable, more expensive model every turn, occasional prediction miscalls, and no clean signal for when to trust the plan versus what's happening live — and they note the harness is "designed to be replaced by the simpler loop" if frontier models become fast enough.
Why it matters: For AI-in-education builders targeting young children, this team's architecture proves sub-second tutoring is achievable without sacrificing frontier-model reasoning, but only by rejecting the standard agent loop and owning generation-execution decoupling, async planning, and parallel safety classification themselves — a maintenance burden the team says their harness is "designed to be replaced" once models get faster.
