Wonder

Inspiration

Kids are the most curious people on the planet — they ask "why is the sky blue?" and "how do fish breathe?" all day long. But the tools that could answer them aren't built for them. Typing-based AI assumes a child can read and write fluently. General chatbots aren't safe for young kids. And almost everything is English-first, which leaves out millions of children who think and speak in another language.

I wanted to build something that meets a child exactly where they are: speaking, in their own language, with answers that are warm, simple, and safe. Something that doesn't just answer a question, but nudges a kid toward the next one — because curiosity grows when it's fed.

What it does

Wonder is a voice-first curiosity companion for children ages 6–12.

A child taps the microphone and asks anything out loud — in English, Hindi, Spanish, or another supported language. Wonder understands the question, generates an age-appropriate answer, and reads it back in a warm, expressive voice. After each answer, three "Wonder More About…" cards appear, branching into related questions that build a learning trail and keep the child exploring. There's also a story mode that retells any topic as a short tale with a gentle lesson.

The whole experience is audio-first and reading-optional, so even a young child who can't type or read fluently can use it independently — and a strict safety layer keeps every answer kid-appropriate.

How we built it

I built Wonder entirely on MeDo, describing the app in plain language and iterating through conversation.

• MeDo generated the full-stack app end to end — I never touched boilerplate or config.
• It set up Supabase for data and storage automatically.
• It built secure backend API routes so my API keys live server-side and never reach the browser.
• It deployed the whole thing to a live public URL with one click.

For the intelligence and voice, I integrated:

• ElevenLabs Speech-to-Text to understand the child's spoken question (with automatic language detection).
• ElevenLabs multilingual Text-to-Speech to read each answer back warmly and expressively.
• A language model to generate age-appropriate answers and the follow-up "wonder more" prompts, guided by a strict system prompt for child safety and tone.

MeDo wired all of these together through its backend, and gives me live analytics on my deployment so I can see how Wonder is actually being used.

Challenges we ran into

Keeping API keys safe. My first instinct was to call ElevenLabs straight from the frontend — which would have exposed my keys to anyone with browser dev tools. I had MeDo move every external API call into secure backend routes so nothing sensitive ships to the browser.

Making AI genuinely safe for kids. A general-purpose model will happily answer questions a 7-year-old shouldn't be asking. Getting the safety behavior right took careful system-prompt design — gently redirecting off-limits topics, calibrating vocabulary to the child's age, and never inventing scary "facts."

Multilingual consistency. The spoken answer, the on-screen text, and the follow-up prompts all needed to stay in the same language the child used — which meant threading language detection cleanly through the entire voice-to-answer-to-voice loop.

Resisting feature creep. The hard discipline was shipping one thing that works beautifully instead of five half-built modes.

Accomplishments that we're proud of

• A fully working, deployed product with a complete voice loop: speak → understand → answer → speak back.
• A UI that genuinely feels like it was made for kids — I described the feeling I wanted (bright, gentle, picture-book warmth) and MeDo nailed the tone from description alone.
• Real multilingual support that works across speech input, the answer, and the spoken output.
• A safety layer I'd actually trust around a child.
• Going from idea to deployed app without writing infrastructure by hand.

What we learned

• No-code can do real full-stack. MeDo handled the backend, database, secure routes, and deployment far faster than building them by hand — the bottleneck became my ideas, not my setup.
• Prompt design is the product. Both the kid-friendly UI and the safe, warm AI behavior came down to describing intent precisely — once for MeDo's generation, and once for the model's system prompt.
• Never trust the frontend with secrets — backend proxying isn't optional once real users (or a viral demo) show up.
• Voice-first design is an accessibility unlock, not just a gimmick — it opens the product to kids who can't yet read or type.

What's next for Wonder

• A parent dashboard showing what their child has been curious about, with weekly highlights.
• More languages, especially under-served regional ones.
• Saved "wonder journeys" so kids can revisit and continue past learning trails.
• Auto-generated illustrations for each answer to make explanations even more vivid.
• A classroom mode for teachers to spark group curiosity.
• A native mobile app for tablets, where this experience really belongs.

Built With

• medo
• supabase