Ensuring accessibility in digital textbooks is a critical challenge, especially for visually impaired students and teachers. In this talk, we present the design and implementation of an AI-powered system that automatically generates alternative text (alt text) for images embedded in EPUB files. This system was developed to support inclusive education by enhancing the accessibility of digital educational content. We will walk through the evolution of the solution: starting from a FastAPI-based prototype capable of extracting, analyzing, and captioning images using deep learning models, to a fully serverless deployment using Magnum, which allowed us to scale on demand, reduce operational overhead and costs. Reliability is a cornerstone of our design. To continuously verify the health and correctness of the service, we built an application-level probe—an automated test harness that simulates real user interactions. This probe is executed daily through Dagster, our orchestration tool of choice, enabling robust observability and alerting. The talk will include architectural diagrams, code snippets, and practical lessons learned from integrating AI solutions, serverless infrastructure, and orchestration tools in a production context. Attendees will gain insights into building accessible-by-design AI services, deploying them scalably, and ensuring their long-term reliability, employing python.