yeah, training the second model in edge impulse will work for you, but only accidently since the output of your first model has 3 classes that you can “pack” into an image as RGB. it wouldn’t be as easy with a different number of classes.
that TFJS model all looks sensible