I need a deep-learning foundation model that can reliably spot Diabetic Retinopathy and Glaucoma in fundus photographs, even when the images come from clinics or cameras it has never seen before. I will rely on publicly available datasets, so part of the job is to help me choose, download, and harmonise the best open-source collections, handle class imbalance, and set up robust cross-dataset evaluation. Once the data pipeline is in place, build and train a modern architecture—CNN, Vision Transformer or a hybrid—optimised for both accuracy and domain generalisation. Strong data-augmentation, colour normalisation and adversarial or contrastive techniques are welcome if they improve out-of-distribution performance. Deliverables • Cleaned and documented dataset splits with the code to reproduce them • Training code (Python, PyTorch or TensorFlow) with clear README • Pre-trained model weights and an inference script that runs from the command line and outputs disease probabilities for each eye image • Evaluation report comparing performance on at least two separate public datasets and a small unseen hold-out set • Brief note on future extensibility to Macular Degeneration I will test the model on my own images; acceptance depends on ≥ 85 % AUC for each disease and no major drop when the source dataset changes.