Our experts began with a thorough evaluation of the customer’s manufacturing processes management, which included interviewing department managers and analyzing the company’s IT infrastructure. Based on our findings, we identified several areas of inefficiency and ranked them according to their business impact, selecting the quality control process as the primary AI implementation target to address the limitations of manual inspections, which were time-consuming, error-prone, and inconsistent.

We conceptualized a computer vision solution powered by neural networks that would process product images captured by existing cameras from production lines to detect different flaws on plywood surfaces. The solution would then classify plywood sheets by quality grade based on the amount, size, and severity of product defects and direct the panels to corresponding storage destinations.

Next, we studied the customer’s current product quality inspection data, which included thousands of plywood panel images, to determine its quality and diversity, confirming that we could use existing images instead of collecting new data to train the AI model.

Next, we labeled 5000+ selected images with different defect attributes, such as number, color, area, and severity, and trained the model so that it could correctly identify flaws and handle edge cases.

After that, we divided the solution’s implementation process into three stages. First, we performed model backtesting to evaluate its operation and accuracy, designating key specialists from the customer’s quality assurance department for data verification. Second, we rolled out the pilot solution to a single production line to assess its results on a smaller scale, making it run in a shadow mode alongside human inspectors performing manual assessments. The selected QA experts analyzed the outcomes of the operating AI solution and compared them with human decisions. Using the detected false positives and negatives, we conducted additional image relabeling and fine-tuned the model, increasing its accuracy rate from 95% to 98.3%.

Finally, we deployed the AI defect detection solution to all 14 production lines and provided user training in the form of on-site workshops. Post-launch, we provided ongoing solution monitoring and support, tracking the solution’s metrics to assess its effectiveness over time and plan its enhancement.