For our client, a Tier-1 electronics manufacturer, a one percent defect rate in PCB assembly meant millions in losses and potential failures in medical and aerospace equipment. Their legacy rule-based inspection produced so many false positives that good boards were scrapped and operators spent their days on unnecessary rework. They engaged us to replace it. We built an AI-powered optical inspection suite that detects sub-millimetre soldering defects and component misplacements in real time, mounted directly on the production line.
View Impact
The client needed inspection that was both faster and more trustworthy than what they had. We delivered custom-trained YOLOv11 models detecting missing components, tombstoning, and bridge-solder shorts at sub-millimetre resolution, processing five boards per second under 200ms latency and signalling the line PLC to reject defective units instantly.
Manual inspection was slow, inconsistent, and vulnerable to fatigue. The legacy rule-based vision system produced high false-positive rates, flagging good boards as defective and driving significant manufacturing waste alongside unnecessary manual rework.
We labelled a specialised dataset of 50,000 PCB images including rare corner-case defects, then used transfer learning from YOLOv11 to reach high accuracy on a relatively small industrial set. A multi-camera rig captures four angles to catch bent pins, models run on NVIDIA Jetson edge devices mounted at the conveyor, and confidence gating routes borderline cases to a supervisor for a one-click audit.
Defect escapes reaching customers fell by 82 percent and manufacturing waste dropped 15 percent, because failing assembly machines are now identified within minutes rather than hours. Inspection runs three times faster than the legacy system, lifting overall factory throughput by 10 percent, with 99.2 percent recall on critical defects.
View Impact
The client needed inspection that was both faster and more trustworthy than what they had. We delivered custom-trained YOLOv11 models detecting missing components, tombstoning, and bridge-solder shorts at sub-millimetre resolution, processing five boards per second under 200ms latency and signalling the line PLC to reject defective units instantly.
Manual inspection was slow, inconsistent, and vulnerable to fatigue. The legacy rule-based vision system produced high false-positive rates, flagging good boards as defective and driving significant manufacturing waste alongside unnecessary manual rework.
We labelled a specialised dataset of 50,000 PCB images including rare corner-case defects, then used transfer learning from YOLOv11 to reach high accuracy on a relatively small industrial set. A multi-camera rig captures four angles to catch bent pins, models run on NVIDIA Jetson edge devices mounted at the conveyor, and confidence gating routes borderline cases to a supervisor for a one-click audit.
Defect escapes reaching customers fell by 82 percent and manufacturing waste dropped 15 percent, because failing assembly machines are now identified within minutes rather than hours. Inspection runs three times faster than the legacy system, lifting overall factory throughput by 10 percent, with 99.2 percent recall on critical defects.
Check these Also
Forging Digital Solutions with precision and passion. From freelance roots to global impact.
Trusted & Verified
