Stefan Ebner
Backend Guru
Latency fears my name
Take a closer look at the development process, the technology stack, and the key features of our project, explained with both text and images.
This large-scale multimodal model forms the core of the image-to-text pipeline. Integrated through Vertex AI, it processes individual images as well as bulk uploads, producing precise alt texts across multiple languages. Carefully engineered prompts and evaluation logic ensure consistent quality aligned with accessibility standards.
Langfuse provides detailed observability for all AI-driven processes. It tracks token usage, model latency, error rates, and operational costs, enabling informed optimisation decisions during development and deployment. This ensures long-term stability and transparent performance metrics.
Lottie animations enhance the usability and aesthetics of the admin interface. Their lightweight, vector-based structure enables smooth transitions and clear visual guidance without compromising performance, contributing to a refined and engaging user experience.
R-Clone supports structured file management within the bulk-processing pipeline. It synchronises datasets, prepares image source collections, and ensures reliable handling of large media volumes before they are passed into the generation service.
Google Cloud Platform hosts the complete application stack, from compute resources to serverless components. Infrastructure design, automated deployments, and operational monitoring ensure a resilient environment capable of handling production workloads. The cloud setup supports scalability, security, and efficient long-term operation.
FastAPI powers the system’s central service layer and exposes the REST endpoints used by the CMS, admin interface, and automation components. It manages communication across the database, batch processor, and LLM pipeline, providing a stable and high-performance integration framework.
Bruno enables systematic verification of endpoint behaviour throughout development. Test collections ensure functional accuracy, validate data flow between services, and support regression testing—resulting in a reliable and maintainable API architecture.
Locust simulates high-volume traffic to assess system behaviour under real-world load. These tests identify performance limits, detect bottlenecks, and verify that the alt-text generation pipeline remains stable even during peak processing tasks.
Firestore serves as the central data repository for image sources, generated alt texts, confidence scores, and multilingual variants. The NoSQL schema is designed for fast access, structured document storage, and seamless integration with the generation and admin services.
Eventarc triggers initiate automated workflows whenever new CSV files are uploaded to Cloud Storage. The system extracts, validates, and forwards all image URLs for processing, enabling efficient handling of large-scale datasets without manual intervention.
The React-based admin dashboard provides a clear and functional interface for reviewing, editing, regenerating, and validating alt texts. It integrates AI testing tools, real-time updates, and Google OAuth authentication to ensure usability and secure access for authorised staff.
Bootstrap supports a consistent design language across the admin interface. Its responsive grid system and styling components ensure accessible layouts, clean visual structure, and seamless user interaction on all device types.