IRIS | Devpost

IRIS

IRIS — Intelligent Retinal Insight System

In many African primary-care settings, a healthcare worker may be the first — and sometimes only , point of contact for a patient with an eye condition. Yet access to ophthalmologists, reliable internet, and advanced diagnostic tools can be limited.

What if an AI assistant could provide useful ophthalmology support directly at the point of care , without needing the internet?

That question inspired IRIS (Intelligent Retinal Insight System).

IRIS is an offline, on-device AI ophthalmology assistant designed to support community health workers with eye-health triage, patient education, and referral guidance. It is designed specifically for resource-constrained environments, where cloud-based AI may not always be practical or reliable.

At its core, IRIS runs the lightweight Qwen2.5-1.5B-Instruct language model locally using llama.cpp. We benchmarked different model sizes and quantization configurations and selected the 1.5B model because it provided a strong balance between clinical knowledge, inference speed, and memory efficiency.

On a CPU-only laptop, IRIS achieved approximately 12 tokens/second, used only 1.7 GB of peak RAM, and achieved 66% accuracy on the mmlu_clinical_knowledge benchmark.

But IRIS is more than just a smaller AI model. It represents a different approach to healthcare AI: instead of assuming that every clinic has high-speed internet, powerful GPUs, or access to specialists, we designed the system around the realities of the environment.

IRIS is built to help healthcare workers understand eye-health concerns, identify cases that may require urgent referral, and communicate information more clearly , while keeping sensitive patient information on the device.

Our long-term vision is to evolve IRIS beyond a text-based assistant into a multimodal, offline ophthalmology platform and offline patient records.

If the internet cannot reach the clinic, the intelligence should.