AI Dynamic Sorting System

The problem

Sorting small hardware parts by hand is slow and easy to get wrong. When screws, nuts, and washers vary in size, shape, and material, an operator has to make hundreds of fast judgement calls an hour, and every misclassification feeds rework, quality issues, and higher labour cost downstream. Keeping that accuracy steady across a full shift, and under changing factory lighting, is the part manual sorting struggles with most.

The solution

Life Value built an automated vision system that classifies each part as it passes the camera and routes it without an operator in the loop. The stack was chosen to stay accurate while running on inexpensive, low-power hardware:

• Computer vision in Python and OpenCV. Each part is captured and pre-processed for stable detection under variable lighting.
• TensorFlow classification model. The model identifies the part type and reaches 98% accuracy across screws, nuts, and washers.
• Raspberry Pi at the edge. Inference runs locally on a single board, so sorting happens in real time with no cloud dependency.
• Adaptive recognition. The model handles parts of different size, shape, and material, and can be retrained as new parts are added.

The results

The system replaced manual classification with a consistent, repeatable process and held its accuracy across changing conditions on the line.

• 98% sorting accuracy. Fewer human errors and misclassifications.
• 30% efficiency increase. Faster workflows and better use of operator time.
• Stable under varied conditions. Reliable results across different part sizes, shapes, and lighting.

How it works

A camera trigger captures each part and passes the image to the processing module, where OpenCV prepares it and the TensorFlow model returns a classification. The Raspberry Pi then drives the matching indicator and routing for screws, nuts, or washers. Because the model and logic run on the board itself, the whole loop stays local and fast, and careful calibration keeps image capture steady even when factory lighting shifts.

FAQ

What parts can the system sort?

It was built to classify small hardware components, screws, nuts, and washers, and the model can be retrained to recognise additional part types.

How accurate is it?

The vision model reaches 98% accuracy on the trained part types, which removed most of the misclassification that manual sorting produced.

Does it need expensive hardware or a cloud connection?

No. Inference runs on a single Raspberry Pi at the edge, so the system works in real time without sending data to the cloud.

Can it cope with changing factory lighting?

Yes. The capture and pre-processing steps were calibrated to keep detection stable under varying lighting on the line.

Conclusion

The AI Dynamic Sorting System shows how machine learning and real-time image processing can replace a manual, error-prone task with a dependable automated one, all on affordable edge hardware. It cuts errors, improves throughput, and adapts as new parts are added. The project reflects how Life Value builds practical, scalable AI that delivers measurable value for engineering and manufacturing teams.