Hello there. I work on the TensorFlow.js team here at Google. Just joined this site so wanted to give a response. Essentially given you are more proficient in Python you may prefer to use that stack but do consider the following edge cases for times where you may need to use them:
1) Speed. TensorFlow.js Node can run faster than TensorFlow Python due to the JIT compiler that JS has. Up to 2x faster for full end to end including pre/post processing. See this blog: https://blog.tensorflow.org/2020/05/how-hugging-face-achieved-2x-performance-boost-question-answering.html
2) If you need to run your model client side then TensorFlow.js is the only option for the browser. By doing that you get the following unique super powers:
a) Privacy - no data needs to be sent to server for inference
b) Cost savings - no expensive server with big GPU needed to run 24/7 to provide a service to your users
c) Can run on device with lower latency - no round trip time to server. Some of our pose models for example run at over 120 FPS on a 1070 GPU in the browser!
d) Reach and scale of the web. Any one anywhere can experience the ML model you make by visiting a link. No Linux setup, no GitHub cloning, no CUDA installation, or deps. It just works. If you want feedback and eyes on your latest model, TensorFlow.js is a great way to showcase it and allow anyone anywhere to try it. You may even find bias etc you need to fix by getting more eyes on it. Very useful!
Good luck on your journey and if you are interested I am launching a brand new course - the first on EdX for TensorFlow.js for Web ML here: https://www.edx.org/course/google-ai-for-javascript-developers-with-tensorflowjs