---

The problem

Children learn values from stories, not from lectures. A child forgets being told to be brave, but remembers the hero who was. Reading together at bedtime is also the one screen-free ritual that reliably brings a parent and a child closer, and it builds the reading habit at the age when it still forms easily.

But finding the right story is hard. A book on the shelf rarely teaches the value your child needs this week, it never features your child, and a parent at the end of a long day has no energy left to invent one.

What Narrly does

A parent picks their child's name, the value they want the story to teach, and — if they like — the people the child loves, who join the tale as characters. Narrly writes an original three-panel adventure where the child is the hero, illustrates it, and narrates it aloud. Under a minute, in any of eight languages.

The narration can be read by a story character, or in the parent's own cloned voice — so a child whose mother or father works away can still fall asleep to their voice.

Inside the app

Features

• The child is the hero. Name, appearance, and a cast of family, friends and pets woven into the plot.
• One value per story. Twenty of them — kindness, honesty, courage, gratitude, patience, respect for elders.
• Illustrated. Three panels, with the same character rendered consistently across all of them.
• Narrated. Character voices, or a parent's voice cloned from a ten-second sample.
• Eight languages, including Kazakh — narrated by a native Kazakh speech model, not an accent approximation.
• A library of world folk tales, free and unlimited, including Kazakh tales such as Aldar Kose.
• A reading path with levels, stars, achievements and collectible profile flair, to keep a child coming back.
• Built for parents, not for engagement. No ads, ever. No autoplay. No infinite feed. Children cannot make purchases.

Architecture

A React Native client talks only to Supabase Edge Functions. Every model provider key lives server-side; the app never sees one. Each AI capability is a chain — if a provider is down, throttled, or out of credit, the next one takes over transparently.

flowchart TD
    A["React Native · Expo<br/>iOS · Android · Web"] -->|"HTTPS + Supabase Auth"| B["Supabase Edge Functions<br/>(Deno)"]

    B --> C["generate-story"]
    B --> D["generate-illustrations"]
    B --> E["generate-audio"]
    B --> F["clone-voice"]
    B --> G["dodo-webhook"]

    C --> C1["DeepSeek V4 Flash<br/>→ Claude 4.5 Sonnet<br/>→ Groq → Gemini"]
    D --> D1["FLUX via fal.ai<br/>text-to-image, then image-to-image<br/>to keep the same child in every panel"]
    E --> E1["Kazakh: FreedomSpeech (native)<br/>→ self-hosted MMS-TTS-kaz<br/>Other languages: Higgs TTS 3"]
    F --> F1["ElevenLabs<br/>instant voice cloning"]

    B --> H[("Postgres<br/>Auth · Storage · RLS")]

Things worth a look if you are reading the code:

• supabase/functions/_shared/guard.ts — one guard for every function: auth, rate limiting, monthly story quota, upstream timeouts, and generic errors so a provider failure never leaks its identity to the client.
• supabase/functions/generate-story/index.ts — the child's name travels to the model as data, never as instructions, so a name like ignore previous instructions is written into the story rather than obeyed. Names are also profanity-checked before they can reach a prompt.
• supabase/functions/generate-illustrations/index.ts — the first panel's illustration becomes the reference image for the next two, which is what keeps the same child on every page.
• supabase/functions/generate-audio/index.ts — long stories are chunked, synthesised in parallel and merged back into a single WAV, then cached so a re-listen costs nothing.

Tech stack

Repository layout

mobile/       React Native + Expo app — iOS, Android and web from one codebase
landing/      Static marketing site (narrly.org), hand-written HTML/CSS, 8 languages
supabase/     Edge Functions (Deno) and SQL migrations
infra/        Self-hosted Kazakh TTS endpoint
ios-launch/   App Store assets

Running it locally

You will need your own API keys — none are included here, and none ever were (see below).

# 1. The app
cd mobile
npm install
cp .env.example .env.local        # fill in your Supabase URL + anon key
npx expo start                    # press w for web, or scan the QR code

# 2. The marketing site — no build step, it is plain HTML
cd landing
python -m http.server 8000

# 3. The backend
npx supabase functions deploy generate-story
npx supabase secrets set DEEPSEEK_API_KEY=... FAL_KEY=... FREEDOMSPEECH_API_KEY=...

Only EXPO_PUBLIC_* values are read by the client. Model provider keys live exclusively in Supabase Edge Function secrets and are never bundled into the app.

What is intentionally not in this repository

This is a public mirror of a live commercial product, published so the work can be read and reviewed. A few things are held back on purpose, and the code says so where they are:

• The story prompts and the value framework. The narrative voice, the three-beat quest structure and the per-value teaching guidance are the craft of the product. Everything around them — the provider fallback chain, input sanitisation, the prompt-injection guard, the JSON contract — is exactly what runs in production.
• The art direction behind the illustration prompts, generalised for the same reason.
• Product and business documents, and every environment value.

No API key has ever been committed to this project. Secrets live in Supabase Edge Function secrets and in a local .env.local that git has never tracked.

Status

Live and in early access. The web app runs at app.narrly.org; Android builds ship through EAS, and iOS is in preparation.

License

© 2026 Narrly. All rights reserved.

The source is published for reading and review. It is not open source: you may not use, copy, modify or distribute it, in whole or in part, without written permission. See LICENSE.

---