The team launched the first real‑time AI tutor for children ages 4‑9, merging educational theory with low‑latency design. The goal was to teach math and reading while keeping a child’s attention, which demands instant feedback and natural dialogue.

A 2‑second pause feels different to a five‑year‑old than to an adult, so the system must deliver a sub‑second response on every turn. Frontier models emit the first token in 2–3 seconds and decode at about 30 tokens per second. A standard tool loop would add 3‑4 seconds of downtime, causing children to lose focus.

To bypass this, the developers built a custom harness that streams multiple actions in a single response. An interpreter parses and executes each action while the model continues generating the next ones, allowing the child to see the first move after roughly 30 tokens in. The asynchronous planner runs concurrently, reviewing conversation against lesson goals without blocking the present interaction.

Safety must gate every action in real time, so a separate classifier runs in 500–1000 ms. Waiting for that check would add an unacceptable delay, so the system pre‑generates potential responses and validates them instantly. The approach trades higher compute cost for the uninterrupted learning loop required for young learners.