← All projectsComputer Vision · Machine Learning2026

WeldSight

Weld defect detection you can interrogate

77.3%: val mAP@0.5
6: built-in test images
0.05: confidence floor sent

“A detector that only prints 'defect: 0.81' asks you to trust it. This one is built the other way round.”

The problem

Weld inspection is manual and inconsistent between inspectors, and a model that returns a single number does nothing to fix that — it just relocates the trust problem. The score is never the product; the evidence is.

The approach

A YOLOv8-L detector finds the faults, and the interface shows every box it drew, how sure it was, and exactly which detections produced the verdict. PASS / REVIEW / REJECT is a pure rule you can read in twenty seconds, not a model output. The server sends every box above 0.05 once, and the threshold slider, class toggles and verdict badge all recompute locally — which is why dragging the slider is instant and the number at 0.42 is the model’s actual output.

What makes it
work.

Live confidence threshold from 0.05 to 0.95 with no round trip and no re-inference.

Boxes rendered on a canvas the app controls: hover to highlight its row, click a row to fly the viewport to that detection, hold to see the untouched photo underneath.

Every verdict states the count and the confidence that produced it — '4 fault regions detected, strongest at 92% confidence.'

Grew out of the IGCAR internship pipeline, rebuilt as an interrogable product.