Promark Electronics

Promark conceived the idea of visually inspecting parts using a camera encased in a 3D printed station. An operator on the production floor would position a component to be inspected, such as a cable, under the camera, which would stream video to a connected device. A sophisticated computer vision model would then analyze the component and determine its acceptability based on various criteria. The AI model would scrutinize key aspects of the part (e.g. whether the screws are properly in place or whether the end connections are correctly executed) to ensure each part meets the manufacturer's strict quality standards.

If the component is deemed “acceptable” it would be recorded in the system as such and delivered to the customer. If the component is deemed “unacceptable,” the operator would then manually inspect the item, either confirming the defect or flagging a false positive in the system. Timestamped photos of each part would be added to the database for complete traceability, should that component ever fail down the line.

Promark envisioned a system with five key sections to optimize their inspection process. The Wizard acts as the interface for Promark production employees to create new inspection points by submitting pictures of parts for AI analysis. This section also trains the AI model to detect issues accurately. Vision utilizes the camera to identify defects in parts, employing AI for precise defect analysis. The Inspector serves as the validation interface for final quality inspectors, enabling them to verify part quality with data integrated from Vision. This ensures rigorous quality control, allowing users to manually review inventory and inspect for false positives among parts. The Client Admin Panel is designed for individual clients to manage inspection points, requests, users, and roles, equipped with a dashboard that offers strategic insights. Lastly, the Super Admin Panel provides overarching system control, model management, and the ability to review feedback and communicate directly, facilitating operations across all clients.

Beyond developing the AI-powered models, a central challenge involved their maintenance and scaling once in production. Whereas the realm of machine learning (ML) focuses on creating models that can learn autonomously, Machine Learning Operations (ML Ops) emphasizes the automation and routine training of models to prevent them from becoming less effective as production variables change over time. Our task was to develop a comprehensive pipeline encompassing data collection, managed by the Wizard, through to automatic ML training facilitated by Google Cloud infrastructure, and culminating in deployment via continuous integration and continuous deployment (CI/CD) processes.