My first pass at getting this up involved reskinning that whack-a-mole game in React to use 10 real and 10 fake Kanji images I sliced up. I then also added scoring functionality to support the notion of a bad whack and a strategic miss - you only earn points by whacking the fakes and by letting real ones go unwhacked. The version my wife played in the morning had limited samples, so two play-throughs and she saw every fake kanji I could throw at her...but she got one wrong! Eureka!

For my next pass, I needed an endless amount of samples to trip her up. I forked SketchRNN and made changes to the drawing functions to guarantee that the real stuff from KanjiVG and the output from the model looked stylistically indistinguishable. This is effectively an API endpoint to generate either type of character depending on the arguments passed, but instead of getting too ahead of myself and throwing it on a server for real-time generation, I just cranked out one thousand characters each and put them on S3. As long as characters are shown at random, I figured that was more than enough to challenge her. Lastly, I fine-tuned the scoring feedback so it was crystal clear when she got something wrong (or right).

As of mid-May 2018, her average score is 85%. She's furious and addicted. She got her mother and friends (all native speakers) to try it out and they haven't done much better. Yup. I'm patting myself on the back. It's impossible to master.