HanLevel v1.0

Analysis result
Banner
Home page
Analysis details
Ai Tutor - Mongle

Inspiration

Korean learners often know a text feels difficult, but not why. I wanted to build a tool that could explain difficulty transparently instead of only translating the text or giving a black-box AI answer.

What it does

HanLevel analyzes Korean text and estimates whether it is Beginner, Intermediate, or Advanced based on vocabulary, grammar/morphology, and sentence length. Then Mongle (몽글), the AI tutor, uses that analysis as context so learners can ask things like:

Why is this text difficult?
What does this grammar mean?
Which words are challenging?
Can you make this easier? The interface supports English, Portuguese, and Spanish.

How I built it

The analyzer uses Python, Streamlit, Kiwi/kiwipiepy, and a compact KRDICT-derived lexical index. The readability score is rule-based: [0.45(\text{Vocabulary}) + 0.35(\text{Grammar}) + 0.20(\text{Sentence Length})] After analysis, the results are passed to Google Gemini as grounding context for Mongle. The key design choice is that the AI does not decide the readability level; the analyzer does. The AI explains it.

Challenges I ran into

My first idea was to repeatedly generate simpler versions until they reached a target score, but it was slow and unreliable. I redesigned the project around a grounded tutor instead. That made the AI more useful, faster, and better aligned with the original purpose of HanLevel.

Accomplishments that I'm proud of

I'mespecially proud of turning HanLevel from a static readability profiler into a more complete learning experience. For v1.0, I added Mongle (몽글), a grounded AI tutor that uses HanLevel’s own analysis as context instead of generating explanations blindly. I also added multilingual support in English, Portuguese, and Spanish, conversational follow-up, and a clearer learner-focused interface.

What I learned

I learned a lot about combining rule-based NLP with generative AI, prompt grounding, multilingual UX, Korean morphological analysis, and designing AI features around what they are actually good at.

What's next for HanLevel v1.0

Next, I want to improve HanLevel with a larger external evaluation set, better context-aware vocabulary disambiguation, and richer pedagogical grammar analysis. I also plan to expand Mongle with stronger grounding evaluation, more learner controls, additional languages, and potentially teacher-facing features.