Runway: Car Image Analysis Model Operating Process Automation

A monolithic architecture, where only one model can be deployed, is difficult to maintain. Changes to even a few modules affect the entire system, reducing operational reliability. The separation of image data collection and management, model development and deployment, operation, and monitoring environments limits the process of retraining to improve the performance of image analysis models. Visualization is needed to effectively monitor model performance degradation, compare model performance, and manage service quality.

The architecture was switched to a parallel microservice model to minimize the impact of module changes on production services and create a scalable deployment environment. Development and production environments were integrated, and retraining pipelines were established to provide sustainable AI services on a unified platform. An intuitive GUI was implemented to visually monitor A/B testing between retrained models and service models.

Initially, only a single model was deployed, but now eight different models are used in a complex manner. The retraining process, previously conducted annually, is now performed frequently on the AI platform (Runway), with models updated quickly to match the new car release cycle. Utilizing a site licensing policy reduced software maintenance costs by 40% compared to traditional perpetual licenses.