Adding Emotion and Intelligence to Transaction History: A Special Financial Experience Built with Image Calendars and Natural Language Search

The MOZ team, winner of 2nd place at KakaoPay's 2nd Hackathon, presented a financial service that transforms ordinary transaction history into special memories. When users leave notes simply by selecting chips, the service generates an image calendar based on them and allows users to easily find desired entries through natural language search.

The service architecture was built on AWS Amplify, S3, RDS, and EC2. In particular, to reduce inefficiencies that can arise when processing large volumes of transaction history, the team adopted an approach of calling the AWS Bedrock API and executing the Tool calling handler directly on the server.

The Nova Canvas model was used for image generation. A decisive factor was its support for long prompts of up to 1024 characters, allowing detailed context to be reflected. In addition, the team developed an image prompt generator that combines transaction data and user information to let the LLM create optimal prompts, ensuring diversity and quality in the images.