AI grading is an interesting exercise, because it shows just how much you need to understand something to build an AI system that can do it well. This is a good example of a classifier problem, where the classifications are the grades of a coin. A standard approach to this is to throw a lot of image data together with ground truth grades (I know what you're thinking) and let the system learn how to classify unknown images. This will work pretty well with wholesome, circulated coins with flat appearances that can be graded pretty much as line drawings and imaged on a flatbed scanner with repeatable results.
Uncirculated coins pose a different challenge. If you look at how people grade them, there is a lot of tipping and twirling of a coin in the light to get a full picture of what it looks like. If you translate that process to a computer, that is an incredible amount of data that needs to be acquired, and acquired without loss due to an insufficient imaging setup. We've all seen countless GTG threads everywhere that have unexpected grades associated with a coin. We are grading these with one image. Sometimes they're good images, sometimes not, but it's always that case that a single image isn't capturing everything about a coin. More nefariously, it's sometimes the case that a single image is hiding specific things about a coin, such as a patch of hairlines, a bad hit, or a scratch. If I'm trying to rip of eBay buyers and I normally use bad (for example, overexposed) pictures of problem coins to do this, knowing there's an AI grader in play, I might decide to try and game that with my photos.
To me, what would be the most impressive is a data acquisition system that could acquire the data necessary to train such a system well and then to actually perform the grading on test coins. From there, the next step would be to see how to optimize the whole process.
Simpler AI tasks related to coins include identification, coarse grained attribution (think Overton varieties, not VAMs), and maybe even AT/NT assessment.