Case studies

Education · South Korea

TOPIK Exam Preparation Platform with AI-Powered Essay Scoring

Built a TOPIK learning and exam-prep system (B2C, B2B, Admin) on web and mobile for Woongjin ThinkBig, with a Generative AI solution for automatic essay scoring that achieves over 80% accuracy compared with teachers.

JavaSpring BootJava EEAWS EC2AWS RDSAWS S3AWS SESAWS CloudWatchAurora MySQLGPTLlama 3.1LLM Prompt Engineering

Project overview

  • Client: Woongjin ThinkBig – a leading Korean early-childhood education company, developing educational and cultural content for children since 1980; it serves all ages, covering reading, learning and online/offline book publishing
  • Country: Korea
  • Domain: Education – TOPIK Korean-language exam preparation
  • Scale: ~67 MM (including Onsite and Offsite)
  • Duration: 6 months
  • Project type: Web + Mobile App, with AI integration
  • Scope of work: Wireframe, SRS, technical solution design, Graphic Design, application development (website for education centres, mobile app for learners), AI essay-scoring development, function test & integration test, UAT support, deployment to Staging & Production, app upload to the Apple App Store & Google Play

Problem / Challenges

  • In 2024, the client wanted to build a system dedicated to TOPIK learners and candidates, comprising: a B2C Module (individual users), a B2B Module (corporate Operators, Teachers and Students), and an Enterprise Management Module (Admin site for WJ ThinkBig).
  • The client had no in-house IT team and IT costs in Korea are high → it needed a vendor with ample, high-quality resources.
  • Product requirements were unclear before kick-off → a flexible development approach and a BA team for business consulting were needed.
  • The system had to reuse the existing LMS → the vendor had to understand the system quickly and propose a low-impact integration solution.
  • There was no automatic AI scoring solution → AI was needed to score the Writing section with high accuracy at a reasonable cost.

Solution

  • Working model: an onsite PM, BA/Comtor and Tech Lead ensured smooth, effective communication; a ticket & task management system shared with the client; technical consulting and testing.
  • Agile process: 2–3-week iterations (requirement & user story analysis → UI/UX updates → development → unit/integration test → build & deploy via CI/CD → bug fixing → partial UAT/QA), daily meetings, final UAT/QA; early and frequent feedback from the client.
  • Graphic Design: a youthful, modern design with UX optimised for mobile (Web and Mobile versions).
  • AI Auto-scoring: prompting techniques applied to the client's dataset (questions, answers, teacher evaluations) to create optimised prompts, fed into LLMs such as GPT, chatntq-ja and Llama3.1 to return scores in the required format; integrated into the online Korean self-study application.

Key features

B2C Portal (individual users)

  • Multilingual, friendly, easy-to-use interface.
  • Placement test to assess level and suggest a suitable learning path.
  • Korean learning by level across 3 skills: Listening, Reading, Writing.
  • Mock tests by level (Listening, Reading, Writing); past TOPIK papers by level and exam session.
  • Progress tracking, vocabulary and grammar review.

B2B Portal (businesses)

  • Create and manage classes for each teacher according to schedules and learner progress.
  • Create and print homework for learners.
  • Help learners review lessons, take mock tests and practise past papers; track learning progress.

Woongjin Admin (Super Admin)

  • User list management.
  • Create and manage lesson and mock-test content; design multilingual lessons.
  • Manage operational content: payment status, announcements, notifications, support contacts.

Technology

  • Language & Framework: Java Spring Boot, Java EE
  • User Application: Web, Mobile App
  • Infra: AWS Cloud (EC2, RDS, S3, SES, Load Balancer, CloudWatch)
  • Database: Aurora MySQL
  • AI: LLM (GPT, chatntq-ja, Llama3.1), prompt engineering

Results

  • Comparing AI scores with teacher evaluations showed the AI scored over 80% of learners' submissions accurately.

Have a similar project?

Contact SSD