Edge AI-Powered Quality Inspection System for Mumbai-Based Electronics OEM
Project Overview
EmbedCrest, an AI development company and computer vision development services provider based in Mumbai, India, engineered a high-speed edge AI quality inspection system for Vyom Electronics, a leading contract electronics manufacturer in Mumbai. The project showcased our embedded solutions and edge AI development services capabilities, deploying NVIDIA Jetson AGX Orin modules running custom deep learning pipelines to inspect PCB assemblies at full production line speed. Our embedded developers designed a cascaded architecture combining YOLO v8 object detection with a specialized solder defect classification network, trained on 85,000 annotated production images and optimized for real time inference using TensorRT INT8 quantization. The TinyML solutions approach achieved 99.4 percent defect detection accuracy across 23 defect categories while maintaining sub-200ms processing time per board. EtherCAT integration enabled closed-loop feedback to the SMT line for automated process correction. As one of the top embedded systems companies in India, EmbedCrest delivered this turnkey embedded systems solution from our Mumbai engineering center, including hardware design, model training, production line integration, and operator training. The system reduced Vyom defect escape rate from 2.8 percent to 0.12 percent while enabling a 83 percent reduction in manual inspection headcount.
Problem Statement
Vyom Electronics, a Mumbai-based contract electronics manufacturer producing PCB assemblies for automotive and industrial clients, was experiencing a 2.8 percent defect escape rate in their SMT production line. Manual visual inspection by 12 operators caught only obvious defects like missing components and tombstoned resistors, while subtle solder joint defects, insufficient paste deposits, and micro-cracks consistently went undetected until field failures. With annual production exceeding 4 million boards, even a small defect escape rate translated to costly warranty claims and customer trust erosion. They needed an AI development company that could deliver computer vision development services integrated directly into their high speed production line.
Engineering Approach
EmbedCrest developed a multi-camera edge AI inspection system using NVIDIA Jetson AGX Orin modules running custom computer vision pipelines. Four high-resolution industrial cameras with telecentric lenses captured top-view and angled images of each PCB at line speed of 1.2 boards per second. A cascaded deep learning architecture combined a YOLO v8 object detector for component presence and placement verification with a custom defect classification network that identified 23 distinct solder joint defect categories including cold solder, bridging, voiding, and insufficient wetting. The system was trained on a curated dataset of 85,000 annotated PCB images collected over six months from Vyom production lines. Our TinyML solutions approach optimized the inference pipeline using TensorRT with INT8 quantization, achieving under 200 milliseconds total processing time per board across all four camera views. Results were fed back to the pick-and-place machine controller via EtherCAT for real time closed-loop process correction, and a custom dashboard built with our software development services displayed live defect trends, pareto charts, and statistical process control data to quality engineers. The entire embedded solution ran air-gapped on the factory network with no cloud dependency for inference.
Measurable Outcomes
2.8% to 0.12%
Defect escape rate
1.2 boards/sec
Inspection speed
99.4%
Defect detection accuracy
12 to 2 operators
Manual inspector reduction
Technologies Used
“EmbedCrest delivered exactly what we needed: an AI inspection system that runs at full production speed and catches defects our best operators would miss. The defect escape rate dropped from 2.8 percent to virtually zero, and our biggest automotive client renewed their contract citing improved quality consistency. The team from Mumbai understood our production environment from day one.”
Prashant Deshmukh
VP Quality, Vyom Electronics
Note: Client details have been anonymized to protect confidentiality. Project outcomes and technical specifications are representative of actual engagements.
Frequently Asked Questions
How does the system handle new PCB designs that were not in the training dataset?
The system uses a two-phase approach for new designs. First, the YOLO v8 component detector is fine-tuned using a small set of 200 to 300 golden sample images from the first production run, which takes approximately 2 hours of training on the Jetson AGX Orin. Second, the solder defect classifier is design-agnostic as it operates on cropped individual joint images, so it generalizes across board designs without retraining. New designs are typically production-ready within a single shift.
What lighting and camera setup ensures consistent image quality?
We use a custom structured lighting dome with four quadrants of independently controlled LED arrays providing diffuse, directional, and dark-field illumination. Telecentric lenses eliminate perspective distortion across the entire field of view, ensuring consistent magnification regardless of component height. The lighting sequence cycles through four illumination patterns per board capture, revealing different defect types: diffuse light for component presence, directional light for solder joint topology, dark-field for surface cracks, and backlighting for through-hole fill verification.
Can this system be deployed at other manufacturing facilities?
Yes, the architecture is modular and designed for replication. The edge computing unit, camera array, and lighting dome are packaged as a self-contained inspection station that mounts directly onto existing SMT conveyor systems. Our embedded software solutions include a calibration wizard that automates camera alignment, focus adjustment, and lighting optimization for each new installation. We are currently deploying identical systems at two additional electronics manufacturing facilities in Pune and Bangalore.
How are false positives managed to avoid unnecessary production line stops?
The system uses a confidence-tiered response strategy. Detections above 95 percent confidence trigger immediate reject and diversion to a quarantine bin. Detections between 80 and 95 percent confidence flag the board for secondary review by the two remaining human inspectors, who can confirm or override the AI decision. All override data is logged and used for weekly model retraining cycles, continuously reducing the false positive rate. After six months of production deployment, the false positive rate stabilized below 0.3 percent.
