Tutor Luna: how an under-15-person team built an AI spoken-English "system course"

In late June, leading edtech company LingoAce launched Tutor Luna, an AI-native English-learning product for children aged 4–12. In its first month, Tutor Luna far exceeded expectations: 60–70 effective spoken turns in a 20-minute lesson, and August revenue more than 10 times July's.

Rather than competing on model size or feature count, Tutor Luna pioneered a new format — the "AI spoken-English system course" — that targets children's reluctance to speak. It lets them converse at low cost and high frequency with an AI teacher rivaling a human tutor.

Unlike recorded courses (input-heavy) or AI speaking partners (random chat), Tutor Luna simulates one-on-one teaching with systematic interaction and output. Its learning engine — built from four capabilities mirroring a human teacher's workflow — generates content, drives dynamic questioning, remembers student state across ability/memory/emotion, and evaluates not just right-or-wrong but completeness and error type, then updates strategy. The loop: recognize ability → generate task → give feedback → model ability → optimize.

Notably, children speak more with the AI teacher than with a human: freed from the feeling of being watched, they read aloud and gesture freely. Tutor Luna measures success by spoken frequency and completion, not vocab count — evidence that AI can deliver a structured class at scale, with interaction density beyond human lessons, and that users keep coming back (D30 active retention reached 98% among hundreds of paid learners observed through 30 days).

The team, kept under 15 people, runs like "a band" — research, algorithm, product, engineering and marketing each solo yet hear the whole song. Supported by LingoAce's resources but operating with internal-startup autonomy, it went from definition to launch to early commercial validation in about eight months.

Tutor Luna is evolving from an English spoken course into a "capability-growth infrastructure," with staged goals L1–L4: structured execution now, dynamic in-class orchestration in 2027, then long-horizon path planning and, ultimately, a lifelong capability agent. Beyond English, it is building Chinese literacy and math products, and since September is entering more markets — Thailand, Vietnam, Japan and Korea first, then North America and Europe — under a model of globally unified core capabilities with lightly localized content and deeply localized go-to-market.