DreamsComeTrue

A proven framework for enterprise-level application development, suitable for RESTful API development and microservices architecture implementation. Efficiently handles authentication/authorization systems through Spring Security and data persistence management through Spring Data JPA.

Built a message queue-based asynchronous processing architecture to manage the long processing times and complex call chains of generative AI services. Through RabbitMQ, we queued tasks between multiple AI services (ChatGPT, Midjourney, LUMA AI) for sequential processing, and implemented retry mechanisms for failed tasks and error handling using dead letter queues. This effectively managed the unstable response times and errors of generative AI, improving system stability and reliability.

Containerized each component of the microservices architecture to ensure consistency between development and production environments, enabling independent scaling and deployment. Used Docker Compose to configure local development environments and run each service (Spring Boot server, RabbitMQ, MySQL, Redis) as containers, simplifying dependency management and deployment processes.

Utilizes ChatGPT API to analyze user-input dream data, separate it into scenes, and select appropriate background music. Generates images visualizing the atmosphere and content of dreams through Midjourney, and finally creates short-form videos through LUMA AI. Stably calls each AI service's API, processes responses, and implements retry logic for failures to produce high-quality results.

Efficiently managed long processing times of generative AI services through a message queue-based architecture using RabbitMQ. Queued tasks between multiple AI services for sequential processing and executed multiple AI tasks simultaneously through parallel processing to reduce overall processing time.

Implemented retry mechanisms to handle unstable responses from generative AI services. Failed tasks are moved to dead letter queues for separate management and monitoring, improving system stability and reliability.

Utilized Redis to cache AI task states and progress for fast retrieval. Designed the system to store task states in a centralized database and allow each service to query necessary state information for processing.

Implemented a JWT-based authentication system using Spring Security and supported social login through Google OAuth and Kakao OAuth. Implemented token generation, validation, and refresh logic, and added a refresh token mechanism to improve user experience.

The message queue-based architecture using RabbitMQ effectively managed the long processing times and unstable responses of generative AI. Significantly improved system stability by implementing retry mechanisms for failed tasks.

Maintained consistency between development and production environments through Docker containerization, and improved maintainability by clearly separating component responsibilities through Spring Boot microservices architecture.

We lacked strategies for cost management and optimization of generative AI services. While using services with high API call costs, we failed to carefully design strategies to minimize unnecessary retries or duplicate calls.

The lack of real-time feedback or progress indicators for the long video generation time resulted in a poor user experience.

Before introducing generative AI services, I will first conduct research comparing price structures of various models and service providers to find the most cost-effective combination in terms of quality.

To convert wait time into part of the experience, I would like to try providing additional features like dream interpretation or emotion analysis during video generation, or implementing functionality for users to explore videos created by others.

To systematically improve the quality of generative AI responses, I will introduce a prompt version control system and build a process to continuously optimize by creating a system to evaluate the quality of results for each prompt.