Smart Sorters Part 4:
What Did We Do with Google AI?

During our search for AI, we discovered several examples of AI in sorting, ranging from separating images of cats and dogs to a child prodigy training AI to identify cars. A Japanese farmer even used Google AI to sort cucumbers, according to Google AI’s blog. Looking at these examples, we felt that Google AI was promising for Sunsort.

We had a few unique obstacles when it came to integrating Google AI with Sunsort. To achieve ten pockets every second, Sunsort must grade the fruit within milliseconds. This meant that Sunsort^ai had to recognize the stem, blooms, size, color, and a myriad of defects in one pass, all while running on our rugged industrial computers. It would have been easier to put the AI on rack-mount servers with powerful processors, but that would have increased the cost of Sunsort^ai and required a dedicated dust-free space to house it.

So instead, we went about optimizing everything to speed it up. We did everything from syncing the spools to the optics, designing a stable housing to take the clearest and brightest picture, and implementing every measure we could think of to increase the AI's performance.

"...we offloaded the training portion of the AI onto Google Cloud, because training the AI required the most computing power."

Also, we offloaded the training portion of the AI onto Google Cloud, because training the AI required the most computing power. With Google Cloud’s near limitless computing power, Sunsort^ai can be trained with thousands upon thousands of images within a reasonable timeframe. This also meant that packhouses didn’t have to recalibrate or retrain the sorters because of changes in fruit or from a bad training image. In the end, we made Sunsort^ai lean, fast, and capable of leveraging Google Cloud’s computing power.

"...learning how to operate a Sunsort^ai usually only takes thirty minutes to an hour."

Next, we set out to create a user interface to complement Sunsort^ai. First, we put in safeguards to prevent user error. Next, we made sorting out defects as simple as checking a box with a mouse click. Finally, we created a simple method to granularly segment each defect. We also included other features such as the ability to save multiple takeout settings between uses and the ability to control all Sunsort^ais from a single master computer. All the functions are intuitively laid out and easy to grasp. As a result, learning how to operate a Sunsort^ai usually only takes thirty minutes to an hour. The interface is more about helping the packhouses in accomplishing their goals than it is about setting it up or troubleshooting.

Nevertheless, just by looking at its user interface, Sunsort^ai is deceptively simple and lightweight. However, it's not because we stripped it down, it’s because we built Sunsort^ai with a truly intelligent AI that can leverage the computing power of Google Cloud when it needs it the most.

Sunsort^ai is a result of our hard work and truly is what intelligent sorter looks like.