TOPIKnow

Inspiration

For millions of aspiring migrant workers, Korean is more than a language—it is a gateway to safer employment and a bigger opportunity in Korea. Yet many EPS-TOPIK learners still rely on static question banks, memorization, and explanations they cannot easily understand in their own language.

I built TOPIKNow to give these learners something closer to a personal Korean coach: mobile-first, practical, multilingual, and available whenever they study.

What it does

TOPIKNow turns EPS-TOPIK preparation into a complete learning journey:

• 40+ realistic mock-exam sets build confidence and exam stamina.
• Interactive lessons transform official textbook material into vocabulary, grammar, listening, speaking, and role-play activities.
• Context-aware AI explains incorrect answers in the learner’s language instead of showing only a score.
• Voice interview practice uses Korean speech recognition, natural text-to-speech, personalized feedback, and model answers.
• A multimodal assistant can read Korean signs, forms, menus, and workplace documents from photos.
• Progress is saved across lessons, practice sessions, and exams.
• Multilingual support makes the experience accessible to learners from 17 EPS-sending countries.

The result is not simply another AI chatbot. It is a structured learning system that connects assessment, instruction, practice, feedback, and progress.

How I built it

TOPIKNow is a mobile-first Next.js and React application backed by Supabase.

OpenAI models power the multimodal learning experiences:

• GPT-4o mini understands Korean text and images.
• GPT-4o mini Transcribe converts learners’ spoken Korean into text.
• GPT-4o mini TTS provides a consistent Korean-speaking interviewer.
• Grounded prompts provide question-specific explanations and interview feedback.
• Structured output validation and graceful fallbacks keep lessons usable when an AI request fails.

The learning interface is designed as a progressive journey rather than a collection of disconnected tools. Learners move from a level test to recommended lessons, realistic practice, review, and measurable progress.

How I used Codex and GPT-5.6

Codex with GPT-5.6 acted as my product-engineering partner throughout the build.

I used Codex to:

• Trace a large existing codebase and map dependencies before changing learner-facing routes.
• Transform textbook content into reusable interactive lesson components.
• Implement speech, image, tutoring, assessment, and progress flows.
• Design grounded prompts and validate structured AI responses.
• Diagnose mobile audio, state persistence, and responsive-layout problems.
• Write focused unit and end-to-end tests.
• Inspect the product in phone-sized browser viewports and iterate on visual details.
• Refactor repeated lesson logic into scalable components that can support additional chapters.

Codex accelerated implementation, but the important product decisions remained deliberate: which learner problem to solve, what textbook material to preserve, how feedback should be presented, and where deterministic behavior was safer than generation.

Challenges I ran into

Keeping AI grounded

Language-learning feedback becomes harmful when it invents grammar rules or ignores the current question. I constrained each request with the exact lesson, question, choices, correct answer, and learner response. Generated exam content is validated before it reaches the learner.

Making Korean voice practice reliable

Recording and playback behave differently across browsers and phones. I added server-side transcription and speech generation, browser fallbacks, loading states, rate limits, and clear recovery paths.

Turning a textbook into an experience

A textbook contains valuable material, but simply displaying PDF pages does not create learning. The challenge was preserving the source while converting it into short, interactive steps with immediate feedback.

Supporting both guests and returning learners

The first experience needed to work without registration while still allowing authenticated learners to save progress. This required careful handling of local state, database state, and interrupted sessions.

Accomplishments that I’m proud of

• Built a working, coherent product rather than a single-feature prototype.
• Combined reading, listening, speaking, visual understanding, and assessment in one learning flow.
• Created reusable lesson architecture that can scale across the curriculum.
• Made AI feedback specific to the learner’s actual question and response.
• Designed the experience for affordable phones and real mobile study conditions.
• Added fallbacks so core learning does not disappear when an AI service is unavailable.

What I learned

The strongest educational use of AI is not unlimited content generation. It is delivering the right explanation, practice, or encouragement at the exact moment a learner needs it.

I also learned that reliability is part of pedagogy. A tutor that occasionally gives impressive answers is less valuable than one that stays grounded, preserves progress, and works consistently on the learner’s phone.

What’s next for TOPIKNow

Next, I plan to expand the interactive textbook coverage, introduce a deeper mastery model for personalized review, strengthen pronunciation feedback, and build academy tools that help teachers see where groups of students are struggling.

The long-term goal is to give every EPS-TOPIK learner access to a excellent Korean coach—regardless of location, income, or native language.

Built With

• codex
• gpt-4o-mini
• gpt-4o-mini-transcribe
• gpt-4o-mini-tts
• gpt-5.6
• javascript
• next.js
• openai-api
• playwright
• postgresql
• react
• supabase
• tailwind-css
• vercel