English Time

The brief was unusually human: teach one six-year-old Urdu speaker to speak English through a daily, roughly 30-minute live conversation with a friendly character, the way a patient tutor would, adapting to her age and mood, not a gamified drill app.

• A young child needs a genuine spoken back-and-forth, not tap-the-answer exercises, which means low-latency speech in and speech out.

• When she gets stuck, the tutor has to explain in Urdu and then guide her back into English, mid-conversation.

• Learning only sticks with continuity: the tutor must remember earlier days, follow up, and schedule vocabulary review.

1. Built a single Next.js app that runs the 'call': mic capture with voice-activity detection, an animated character that lip-syncs to the reply, and on-screen flashcards, pictures and stars.

2. Had the language model reply with structured JSON describing both what to say and what to show, so the agent genuinely drives the lesson: pick a flashcard, quiz a word, award a star, clear the screen for free chat.

3. Layered spaced repetition to schedule word reviews, and folded an end-of-session summary into a long-term profile for cross-day memory.

4. Kept it self-hostable on a single VPS so a family can actually run it.

Feels like a real conversation

Speech in, speech out, with no noticeable lag, so a six-year-old just talks instead of tapping through menus.

Picks up where it left off

The tutor remembers past sessions and schedules review, so learning compounds day to day instead of resetting every time.

Simple enough to actually run

It's self-hosted on one machine, no complicated infrastructure a family would need help maintaining.

The number that matters here is the streak. Sixty consecutive days is not something you can make a six-year-old do; she keeps coming back because it feels like talking to someone rather than being tested. More than 250 words have stuck, scheduled and re-tested by the spaced-repetition layer rather than crammed.