Fluvia: Practice English AI

Speak English out loud with an AI partner who lets you fix your own mistakes, remembers your weak spots, and gets you talking instead of tapping.

Comment

Inspiration

I'm a software developer in Ecuador, and I started learning English to reach better jobs abroad. The popular AI tutors I tried built my whole study plan around the role I gave them, so I got very good at talking about bugs and debugging and froze on everything else. Most had no accents to choose from and no place to keep the vocabulary I was learning. General chatbots didn't remember anything and weren't personalized at all.

What finally worked was friction: being pushed into topics I didn't know how to talk about. So I built a small local website just for myself, with varied topics and subtopics, different accents, a vocabulary cycle and connected-speech listening. It got me fluent faster than anything else I'd tried. Fluvia is that tool, rebuilt as a product: a hub for mastering spoken English through real conversations, on any topic.

What it does

Fluvia is a spoken-English coach for iPhone and the web.

  • Voice conversations with AI characters. Role-plays with a persona, friendly debates, or topic threads. Each character is an animated avatar whose lips move with its voice.
  • Self-correction. When you repeat a mistake, Fluvia doesn't hand you the answer. It quotes your phrase and asks you to fix it, with a reveal button if you're stuck. When you manage it, you get a visible "you fixed it" moment.
  • Pattern Radar. Home shows the categories of mistakes you're working on, and each category is practised with its own strategy (see Design).
  • Support during a conversation: a hint card, adjustable speech speed, an inline translator, a vocabulary focus sheet, and a spoken review at the end.
  • Free study tools: connected-speech decoding, listening, writing comparison, vocabulary, phrasal verbs, idioms, translation, goals and progress.
  • Five interface languages (English, Spanish, Portuguese, French and Arabic), with full right-to-left layout for Arabic, including the avatar.
  • Try before signing up. New users have a real conversation before creating an account.

How we built it

  • App: Expo / React Native (Expo Router, Reanimated, React Compiler).
  • Backend: Next.js + PostgreSQL.
  • Purchases: RevenueCat (react-native-purchases) for subscriptions. A RevenueCat webhook to our server is the single source of truth for Premium.
  • Engagement: OneSignal for push, in-app messages and Journeys. EAS Update ships fixes over the air.
  • Voice: speech recognition for input and OpenAI text-to-speech for voices. The avatar runtime is shared between the app and the web.

Challenges

  • Keeping voice latency low enough that a conversation feels like one. We measured and removed a speech provider whose delay reached 10 seconds.
  • Knowing what our learners actually get wrong. Our first error tagger (one small model doing six jobs in one call) agreed with a reference labelling only 0.31 (Cohen's kappa). We rebuilt it on real data and reached 0.69 (see Design).
  • Selling from Ecuador. Stripe doesn't onboard Ecuadorian businesses, so iOS runs on RevenueCat and the web on Lemon Squeezy. Both write to the same server-side entitlement.
  • Getting right-to-left Arabic right across the whole UI and the avatar.

What we're proud of

  • Self-correction as a core mechanic, not a gimmick.
  • An error taxonomy built from our own learners' data, with a different correction strategy per category.
  • A free trial sized by purchasing power, running as a live A/B test.
  • A paywall built from the learner's own progress.
  • Pushes that bring learners back with their own words, and never interrupt a live conversation.

What's next

  • Accent-specific voices.
  • Workplace-English scenario packs.
  • Android release after Google Play billing is verified.

Built With

Share this project:

Updates

Submission history