VerseVoyage

Inspiration

Poetry can give children a passport into another language, place, and way of seeing. Yet classic poems are often presented as static text, adult-level translation, or right-or-wrong quizzes. We wanted to preserve the original voice while making sound, imagery, vocabulary, emotion, and cultural context welcoming to young learners.

What it does

VerseVoyage is a free, privacy-conscious poetry journey for children ages 2–15. It turns 111 public-domain poems across 17 languages and 19 countries into interactive reading, listening, Chinese character tracing, missing-line games, age-aware AI conversations, and child-authored poetry quests.

A learner can read an original beside a clear learning translation, listen with language-aware speech, explore cultural context and the arts, ask Lumi an open-ended question, practise a line privately, trace Chinese characters for friendly star rewards, restore a missing authentic line, complete a bounded writing quest, and save progress in a guardian-owned learner profile.

The anthology combines the complete 75-poem Chinese primary-school curriculum collection with 36 world classics.

How we built it

The responsive frontend uses React 19, Redux Toolkit, vinext/Vite, OpenAI Sites, Drizzle ORM, and Sites D1. A Flask service exposes recommendation, tutoring, missing-line challenge, and poetry-quest endpoints. It uses OpenAI moderation and the Responses API with GPT-5.6 Sol. Shared poem data keeps the frontend, backend, and tests aligned.

The recommender is a transparent contextual-bandit prototype. Opens, saves, completions, voice practice, handwriting, and recall games reinforce relevant languages, countries, traditions, and themes; a deterministic exploration bonus prevents the experience from collapsing into a narrow engagement feed.

How GPT-5.6 is used

GPT-5.6 powers three bounded learning experiences:

1. Socratic interpretation: Lumi receives the selected poem, reviewed context, learner age, and question, then returns a concise evidence-based interpretation and a follow-up question.
2. Interactive recall: the correct line remains authentic and curated; GPT-5.6 creates three plausible but incorrect choices, an imagery-based clue, and an age-appropriate explanation.
3. Creative scaffolding: Poetry Studio turns a theme, feeling, and poetic tradition into a starter, word basket, and constraint-based quest while leaving the poem unwritten.

Inputs are bounded and moderated. When the external service or API key is unavailable, the interface clearly labels and uses deterministic educational examples so judges can still complete the flow.

How Codex accelerated the build

Codex was the primary build environment. It helped turn a broad global-poetry idea into a focused learning loop; make product, safety, and architecture decisions; implement the React/Redux, Flask, D1, and recommendation layers; expand and validate the anthology; create tracing and game mechanics; write child-safe GPT-5.6 prompts; build responsive and accessible UI; generate migrations; test the release; and deploy the final OpenAI Sites version.

The commit history documents the full app, learner profiles, curriculum anthology, multilingual voice modes, Chinese tracing, star rewards, and AI missing-line challenges.

Challenges

The hardest design problem was adding game mechanics without turning poetry into a single-answer subject. We kept interpretation open-ended while using exact-answer mechanics only for source-grounded recall and handwriting practice. We also built clearly labeled deterministic fallbacks so the demo never becomes a broken state.

Accomplishments

• A source-aware anthology of 111 public-domain poems.
• An AI learning companion that invites further noticing instead of declaring one correct interpretation.
• Direct Chinese tracing with transparent, encouraging feedback.
• Authentic-line recall games where GPT-5.6 enriches learning without changing the source poem.
• Guardian-owned profiles, no child email, optional microphone access, and no uploaded recordings.
• A deployable release with deterministic demo paths and automated frontend/backend checks.

What we learned

AI is most useful in poetry education when it creates better invitations to notice, compare, and try—not when it produces the final interpretation or poem. Graceful degradation is also part of product design: a transparent fallback preserves trust and makes a demo testable.

What's next

Next we would run educator and cultural-expert review of every learning translation, conduct accessibility sessions with children and teachers, evaluate whether recommendations improve breadth and learning outcomes, and add formal COPPA/GDPR-K review, verified parental consent, deletion/export tools, retention controls, and production abuse monitoring.

Built With

• cloudflare-d1
• codex
• drizzle-orm
• flask
• gpt-5.6-sol
• html
• openai-moderation
• openai-responses-api
• openai-sites
• python
• react-19
• redux-toolkit
• typescript
• vinext
• vite
• web-speech-api