ReadFIT Journeys

Inspiration

ReadFIT was inspired by my friend Bünyamin (“Ben”), a former journalist rebuilding his life in Germany. He studied German for 7-10 hours a day—often past midnight, even while visiting me in Berlin. His extraordinary commitment became my deepest motivation.

I was also inspired by exceptional editorial design, especially The New York Times, whose visual storytelling keeps me reading to the end.

As a Berlin-based tech lead with a love for languages and linguistics, I built ReadFIT around a simple idea: learn through reading what genuinely interests you. After testing it with learners and Reddit communities, Codex and the GPT-5.6 API enabled me to realize that vision fully with ReadFIT Journeys.

What it does

ReadFIT Journeys turns any topic you're curious about into a structured, adaptive learning journey in the language you're learning. You tell it what you want to read; it plans a short course of leveled "stories" at your CEFR level, each one an editorial-quality reading with tap-to-translate on every word, grammar taught in context, a quiz built from the words you looked up, and difficulty that adapts to your feedback. And maybe most critically, it doesn't just present information nicely — the writing, the vocabulary support, and need-based visuals (a map, a chart, a diagram, only when they truly aid understanding) are designed to pull you forward, unit after unit. Learning a language becomes something you actually want to keep doing.

How we built it

ReadFIT began with 8 months of experience building for real language learners. That earlier product gave me domain knowledge, user feedback, and a clearer understanding of where conventional reading tools fall short. For Build Week, I carried that knowledge forward—but created the submitted application as a new codebase during the competition.

I used Codex as my primary engineering partner throughout development. Before implementation, I pressure-tested the product scope and converted the resulting decisions into focused build prompts. I remained responsible for the product direction, learning design, architecture, scope, and final quality decisions, while Codex accelerated implementation, debugging, refactoring, testing, and documentation.

The learning engine uses GPT-5.6 with Structured Outputs and strict JSON schemas:

• The journey-generation pipeline evaluates the learner’s selected language, level, and interests, then creates a structured learning journey containing between one and eight units.
• Each unit is generated independently and includes a level-appropriate reading, contextual word explanations, dictionary forms, grammar guidance, highlighted examples, and a personalized quiz.
• A separate GPT-5.6 storyboard step analyzes the completed unit and determines whether a map, chart, or diagram would materially improve comprehension. It deliberately returns no visual when one would only be decorative.
• After every unit, learners can report that the material was “too easy,” “just right,” or “lost me.” This feedback changes the density, scaffolding, and challenge of the following unit without incorrectly treating one response as a new CEFR assessment.
• Units are generated only when requested, allowing the learner to start immediately without generating an entire journey in advance.

Codex helped implement the complete competition codebase: the Next.js application, Supabase data model, server-side generation routes, interactive reader, journey map, quizzes, adaptive progression, visual-generation pipeline, responsive theme system, usage safeguards, and deterministic Playwright end-to-end tests.

The most important decisions remained human ones: defining the learner problem, limiting the scope to one coherent learning journey, deciding how feedback should affect later units, separating meaningful educational visuals from decoration, and repeatedly removing features that did not strengthen the core experience.

Challenges we ran into

Reliable CEFR leveling is the hard problem in this space — LLMs are notoriously inconsistent at hitting a target level. I'd already learned to solve it not by trusting a label but by generating editorial-quality, structured pieces with validation — and, crucially, by letting the journey adapt to the learner's actual performance instead of the label. To push further for this build, I commissioned a deep-research pass on the best reader and multimodal methods and fed the findings straight back into GPT-5.6's prompts and our architecture: text-first streaming, retrieval-first visuals, and a strict "coherence over decoration" rule for when GPT-5.6 is allowed to add a map or chart. The other real challenge was richness vs. latency — solved by keeping generated text on the critical path and moving every visual into asynchronous, non-blocking generation with prefetching.

Accomplishments that we're proud of

A product in a week. Chat to adaptive journey to leveled reading with tap-to-translate, grammar in context, personalized quizzes, adaptive difficulty, and need-based visuals, in any language and German-first — shipped live with real tests and production safeguards. And a genuinely disciplined AI-assisted workflow.

What we learned

I learned that keeping a “brutally honest” critic in the loop is essential—for challenging assumptions, cutting weak features, and protecting the core experience. Also learned that learners rarely stay engaged with disconnected, generic texts. Reading becomes compelling when it forms a coherent course: shaped by genuine interests, structured around visible progress, and adapted to the learner’s needs. Finally, AI is most effective when it accelerates execution while clear human judgment continues to guide the product, pedagogy, and quality bar.

What’s next for ReadFIT Journeys

The next step is to extend Journeys beyond reading into a more complete language-learning experience. Journeys will also be integrated into the existing ReadFIT iOS app and connected with its vocabulary-practice system, so words discovered while reading can be reviewed and mastered later.

Because AI generation and media processing create real ongoing costs, Journeys will likely be offered as part of a new Pro+ membership with clear usage allowances.

And if ReadFIT wins the competition, the prize will accelerate development, user testing, content expansion, and a dedicated YouTube launch campaign.

Built With

• codex
• gpt-5.6
• leaflet.js
• mermaid
• next.js
• node.js
• penai-api
• playwright
• react
• render
• supabase
• tailwind-css
• ypescript