Multi-Model Interactive Playground

Upload any custom infrastructure photo and enter optional resident notes to evaluate predictions across all 6 model architectures simultaneously.

Holdout Dataset Benchmarks Custom Image Playground

Model Architecture Family 5 Architectures + Ensemble

Tested on 241 Unseen Holdout Samples

ConvNeXt-Tiny (Modern Pure CNN)

Production Winner (Best Real-World Generalization)

92.4% 28.5ms

Calibrated Weighted Consensus (Optimal F1-Soft Vote)

Multi-Model Ensemble

89.0% 28.0ms

Swin-Transformer (Self-Attention)

Best Context

91.7% 34.2ms

INT8 Quantized Dynamic Engine

Fastest CPU

84.2% 12.4ms

EfficientNet-B0 (Baseline Pure Vision)

Baseline

84.2% 32.1ms

Upload Test Sample

Select any local image or drop a photo to evaluate.

Defect Image File *

Click to browse or drag and drop photo

JPG, PNG, WEBP supported

Problem Description (Optional) Enables Multi-Modal Cross-Attention

Quick Prompt Presets:

Awaiting Sample Input

Upload a photo from the left panel to trigger side-by-side inference across ConvNeXt-Tiny, Multi-Modal Bi-Encoder, Swin-T, and INT8 Quantized engines.