AI Diabetic Retinopathy Screener

I need an end-to-end AI solution that can look at single-field retinal photos captured on a handheld smartphone eye camera and tell me, in plain language, whether diabetic retinopathy might be present. The output I want to see is a clear risk flag—low, medium, or high—so a health worker in a rural setting can decide on the spot whether the patient needs referral.

What you will build
• A lightweight model (or API) that ingests JPEG/PNG images straight from the phone’s camera app.
• Logic that classifies each image into low, medium, or high risk for diabetic retinopathy; no further fine-grained grading is required.
• A simple front-end (mobile or web) that lets a non-technical user snap or upload a photo, waits a few seconds, and then shows the risk result.
• Brief user guidance on acceptable image quality (focus, illumination), since capture conditions in rural clinics are variable.

Technical notes
– You are free to leverage transfer learning on public ophthalmology datasets, but the final model must generalise to smartphone-based images.
– If you use cloud services (e.g., TensorFlow Serving, PyTorch, AWS, GCP), make sure the solution can still run offline or in low-bandwidth environments by caching the model locally.
– Accuracy is more important than speed, yet a sub-10-second turnaround on typical mid-range Android hardware is the target.

Acceptance criteria
1. A demo app or web page running the model locally or via a lightweight API.
2. Validation report showing sensitivity, specificity, and ROC curve on an unseen test set of at least 200 smartphone images.
3. Deployment instructions so a rural clinic with limited IT support can reproduce the setup.
4. Source code and pre-trained weights.

If you have already trained models in ophthalmology or other medical imaging, mention it in your proposal; it will speed up shortlisting.