HanaTalk

Inspiration

Hung wants to practice English interviews to apply for jobs in the international market. He wants to practice, but there is currently no English-learning app that allows him to customize a simulated speaking environment tailored to such a specific context. Therefore, Hung decided to use ChatGPT to practice English interviews.

Although ChatGPT allows Hung to upload key data such as his experience, strengths, weaknesses, specific job descriptions, and public company materials to personalize interview practice, its UI/UX does not fully support his daily English speaking practice.

==> PRODUCT IDEA: A personal English-learning assistant that enables Hung to automatically generate personal vocabulary flashcard; highly specific, real-life practice scenarios - improving communication skills based on his own data, powered by AI chatbot and voicebot technology.

What it does

Our product leverages user data including learning goals, proficiency level, and real-life usage contexts to generate fully customized learning experiences.

Instead of one-size-fits-all content, each user gets lessons that are directly aligned with how they actually use the language in their daily work and life.

Key features include:

Onboarding: HanaTalk collects each learner’s target language, current level, goals, and real-world use cases to create practice situations that are highly relevant to them. For professional English scenarios, it can also gather deeper context such as company, role, seniority, and interview or meeting goals.
Vocabulary generation: HanaTalk creates vocabulary flashcard sets based on the learner’s scenario and level. Users can also request custom vocabulary around specific words, topics, or situations they want to learn. In addition, they can provide external materials such as websites, documents, or YouTube links, and the system extracts relevant content to generate contextual vocabulary.
Flashcard learning: Learners receive structured flashcard sets with useful words, phrases, meanings, and contextual examples, making the vocabulary immediately applicable to the situations they care about.
Role-play: HanaTalk generates personalized AI role-play sessions with realistic characters, scenario descriptions, and conversational goals so users can practice speaking English in simulated real-world environments.
Review: After practice sessions, HanaTalk can run review activities to check vocabulary retention, revisit pronunciation and phrase-usage mistakes, and provide speaking feedback across Fluency & Coherence, Lexical Resource, Grammar, and Pronunciation.
Co-create: Learners can co-design additional learning experiences by asking HanaTalk to create new scenarios, flashcards, or practice content from their own topics and source materials.

How we built it

Step 1 - Define user scenarios: We first identified the real-life situations where users need English most, such as job interviews, workplace conversations, presentations, and networking.
Step 2 - Gather rich learner context: We designed an onboarding flow to collect each learner’s level, goals, and use cases, then added deeper context gathering for professional scenarios like company, role, seniority, and communication purpose.
Step 3 -Enrich learning content with external knowledge: We built a research layer that can pull in relevant web results, YouTube videos, transcripts, and user-provided links, then turn that content into usable learning material.
Step 4 - Generate personalized learning activities: Using the learner context and research results, HanaTalk generates flashcard sets, role-play sessions, and review activities tailored to each learner’s needs.
Step 5 - Deliver real-time interactive practice: We implemented a real-time agent workflow over WebSocket so users can interact with the system conversationally, receive structured learning content, and move smoothly from onboarding to practice and review.
Step 6 - Monitor progress and provide feedback: We added a review flow that checks vocabulary recall, revisits speaking mistakes, and updates speaking-performance metrics to help learners see progress over time.

Challenges we ran into

Balancing speed and quality in real-time interaction. Many state-of-the-art models produce better outputs, but their latency is still too high for natural streaming conversations, especially in a speaking-focused learning experience.
Keeping AI conversations natural while still enforcing structured learning flows such as onboarding, flashcard generation, role-play creation, and review. We needed the experience to feel conversational, while still guiding users through a clear learning journey.
Designing truly useful domain-specific learning content. Choosing the right flashcards, phrases, and expressions for each scenario required more than just generic vocabulary generation. We found that relying too heavily on external knowledge can lead to content that is broad or informative, but not specialized enough for practical language learning in a given context.

Accomplishments that we're proud of

Creating a pipeline that can transform external resources into personalized flashcards and practice scenarios.
Designing role-play sessions that feel contextual and purpose-driven instead of generic conversation practice.
Closing the loop with a review system that evaluates retention and speaking performance, not just content generation.

What we learned

Personalization becomes much more valuable when it is tied to real user scenarios, not just language level.
External content is powerful for learning (especially for searching and understanding external factor such as company), but it needs summarization and filtering before it becomes useful practice material.

What's next for HanaTalk

Improve real-time performance by optimizing the agent pipeline and exploring faster model configurations so conversations feel more natural and responsive in streaming mode.
Make the learning experience feel even more human by improving how the system balances open conversation with structured learning goals across onboarding, flashcards, role-play, and review.
Build stronger domain-specific learning intelligence so HanaTalk can generate better flashcards, phrases, and scenario content tailored to each use case, instead of relying too heavily on broad external sources.
Expand the system’s ability to adapt content based on the learner’s context, mistakes, and progress, so each practice session becomes more personalized over time.
Continue refining the quality of scenario-based English learning, especially for professional contexts such as interviews, meetings, and workplace communication.