APPLIED AI · SENIOR DESIGN

OcuScreen↗

Making model output easier to understand.

/04

MY ROLE

Developer · Retinauts student team, UT Arlington

STATUS

Research prototype

TECHNOLOGIES

Python · PyTorch · FastAPI · Next.js

ACTUAL APPLICATION CAPTURE

Explore source code Open prototype

PROOF OF WORK

Open it.
Inspect the work.

Live prototype · 6 tests passed

Explore the working screening interface, the image-quality checks, and an original sprint-report page documenting my individual work.

Open screening prototype Inspect quality-gate tests Inspect inference API Read test run

**My original Sprint 4 report**Page 3 of Adeyemi_Sprint4_Report_UPDATED.pdf, dated September 12, 2026. My self-reported tasks include model experiments, validation, service integration, workflow testing, and demo preparation. Team totals are labeled separately.

tests/test_quality_gate.pyORIGINAL EXCERPT

def test_rejects_small_image():
    result = assess_quality(Image.new("RGB", (100, 100)))
    assert not result.accepted


def test_rejects_blurred_image():
    result = assess_quality(Image.new("RGB", (512, 512), "gray"))
    assert not result.accepted
    assert "blurred" in result.reason


def test_accepts_sharp_image():
    array = np.indices((512, 512)).sum(axis=0) % 2 * 255
    image = Image.fromarray(np.uint8(array)).convert("RGB")
    assert assess_quality(image).accepted

Actual tests reject small or blurred inputs and accept a synthetic sharp image. These checks verify software behavior, not clinical accuracy.

6 local tests passed on September 14, 2026: quality gate, ensemble probabilities, and Grad-CAM hotspot extraction. Research prototype; no clinical validation claim.

01 / THE PROBLEM

A real problem.
A considered solution.

An image classification output is not enough for a useful screening interface. Users also need quality checks, confidence, and a way to inspect what influenced the model.

02 / THE WORK

What I worked on.

Our senior-design project pairs a retinal-image workflow with a Python inference service. It presents five-level severity output, confidence, Grad-CAM visualization, browser-local history, and PDF export.

INTERACTIVE WORKFLOW EXPLORERILLUSTRATED SYSTEM LOGIC

Image input

A retinal image enters the research workflow without patient identifiers.

A guided explanation of the project architecture. This explorer does not connect to production systems.

03 / UNDER THE HOOD

Connect the interface to inference.

A Next.js frontend communicates with a FastAPI service running a PyTorch EfficientNet model.

Show evidence with the prediction.

Quality gates and Grad-CAM attention overlays provide context for the severity output rather than presenting a number alone.

Keep the scope explicit.

The model card records evaluation caveats and intended use. The prototype does not collect patient identifiers.

04 / THE OUTCOME

An end-to-end research prototype with public frontend, inference API, training experiments, tests, and model documentation.

Academic decision-support prototype, not a diagnostic device. The live frontend may require a separately running inference service.

KEEP EXPLORING

Inbox to action

OPEN TO FULL-TIME OPPORTUNITIES

Let’s build
what’s next.

moyosoreadeyemi.dev@gmail.com

Arlington, TX · Open to Seattle and US relocation