UI Design Audit AI v2.1
Summary
A screenshot-first multi-label classifier for detecting 12 UI/UX design-risk categories from mobile and web interface screenshots.
Labels
small_touch_targets, low_contrast, action_overload, navigation_overload, form_friction, content_density, responsive_risk, modal_overuse, deep_scrolling, weak_hierarchy, interaction_overload, mobile_web_mismatch.
Architecture
ExtraTrees multi-output classifier over an 831-dimensional deterministic visual descriptor. Features include global contrast/color/edge statistics, low-resolution intensity/edge/color signatures, row/column projections, grid summaries, and platform signals.
Evaluation
Held-out test set contains 300 examples generated independently from the same deterministic rendering system.
- Micro F1: 0.9411
- Macro F1: 0.9292
- Exact subset accuracy: 0.6200
Per-label results are stored in metrics.json.
Intended use
Design-review assistance, research, prototyping, dataset experimentation, and educational UX analysis.
Limitations
The benchmark uses synthetic screenshots and synthetic labels. Results therefore measure performance on a controlled benchmark, not validated performance on human-designed products. The model should not be used as the sole basis for accessibility compliance, production release, or automated rejection of designs.
Usage
from inference import UIAuditModel
model = UIAuditModel("ui_design_audit_vision.joblib")
result = model.predict("my-ui.png", platform="mobile")
print(result["issues"])