A reading tutor that listens.
A free, voice-first AI reading tutor for kids ages 5–9 — grounded in the Science of Reading, navigable without reading skills, and adaptive enough to slow down, speed up, or scaffold a child in real time.
The problem
Children who are still learning to read can't navigate an app that requires reading. That sentence is obvious — and almost every reading app on the market gets it wrong. They lean on big buttons with labels, modal dialogs with copy, and reward screens with text the child can't parse. The result is that the parent ends up reading the app to the child, which defeats the entire point.
The Science of Reading literature is also unambiguous: kids learn fastest with frequent, short, instruction-led practice that matches their current skill level — and adjusts in real time when they stumble. That kind of personalized loop is exactly what a one-on-one tutor delivers, and it's exactly what no software has been able to do at any scale.
What we built
Lumi is voice-first from the very first interaction. The app opens, an animated firefly mascot greets the child by name and asks what they want to read about today. Every navigation decision the child makes — pick a story, try a word again, take a break — is made out loud. There are no menus. There is no text the child needs to read in order to use the product.
Under the hood, a Claude Sonnet 4.6 session orchestrates each practice round. The session tracks where the child is in the curriculum, what they've struggled with in the last few minutes, and what their long-term progress looks like. When a child reads a word out loud, the audio goes through a speech-to-text pass tuned for child voices, then to Claude — which decides in real time whether to advance the difficulty, slow down, scaffold (sound it out together), or pivot to a different word entirely. The reply gets synthesized in the mascot's voice and played back in under 1.2 seconds. The full round-trip feels conversational, not transactional.
Longitudinal memory is the third pillar. Every session stores a structured summary — words mastered, phonemes still wobbly, what motivated the child today, what bored them. Next session starts with that context loaded, so the tutor picks up exactly where it left off.
My 6-year-old asks for "the firefly" the way she asks for a snack. She doesn't know it's teaching her — she just knows it listens.
How it shipped
Four weeks of pedagogy research and parent interviews to nail the curriculum spine — what a real Science-of-Reading sequence looks like across the 5-to-9 age range. Six weeks of voice plumbing, which was the hardest part of the build by far: child speech recognition is much noisier than adult, latency is critical, and the response model has to handle interruption gracefully because kids interrupt constantly. Four weeks of Claude orchestration — building the per-session memory shape, the scaffolding logic, and the difficulty adjustment. Two weeks of supervised pilot with twenty families before opening to a hundred-twenty for the broader beta.
Free for families is non-negotiable: this is a product the studio cares about for reasons beyond unit economics. A paid school/district tier is in design — same product, plus a teacher dashboard, plus FERPA-compliant data export.