Biggest regret was not keeping a pristine dataset for final testing / evaluation on device. I ended up flying blind when it came to setting the final threshold,

Biggest regret was not keeping a pristine dataset for final testing / evaluation on device. I ended up flying blind when it came to setting the final threshold, testing the effects of quantization, or even just measuring the distortion introduced by cameraphones (compared to directly feeding images during training). What I’m most proud of is the remote neural network injection [0] — which I’m surprised no one has commented on here. I just think it’s absolutely huge to be able to set large parts of your app’s behavior in TensorFlow code, and replace that on the fly in your user’s app as needed. But maybe I’m the only one excited about this :D [0]: https://medium.com/@timanglade/how-hbos-silicon-valley-built-not-hotdog-with-mobile-tensorflow-keras-react-native-ef03260747f3 https://medium.com/@timanglade/how-hbos-silicon-valley-built...