Hack’cessibility
Team consisting of five EPITA AI students ready to build in C++/CUDA (GPU imaging, CNRS), PyTorch/TensorFlow, MLOps/Docker, recommender systems, and medical imaging.
YouTube Video
Project Description
Our project empowers visually impaired people by allowing them to upload photos and receive real-time descriptions of their surroundings. It also supports dual input, allowing users to include both an image and a text query (e.g., “Where is the exit in this room?”) for precise, goal-oriented assistance.
At its core, a multi-agent architecture generates, validates, and refines each image description for accessibility. The refinement loop ensures responses are concise, spatially accurate, and free of redundant or purely visual phrasing, optimized for blind and low-vision users. This iterative process delivers natural, human-like narration that enhances everyday navigation and awareness.
The system runs on Google Cloud Run using NVIDIA L4 GPUs. It uses Gemma 3, a multimodal vision-language model, integrated with a Python FastAPI backend and Docker for containerization. Batch inference, asynchronous processing, and autoscaling on Cloud Run optimize performance and reproducibility.
The goal of our project is to demonstrate how AI can meaningfully enhance accessibility and improve people’s everyday lives through inclusive design.