Research

Can wardrobe classification run entirely on device?

A working investigation into whether practical wardrobe classification can remain on-device without sacrificing the product experience.

This is a working research note, not a claim that one architecture wins in every case. I am evaluating where on-device inference meaningfully improves a wardrobe product and where cloud infrastructure remains the more practical choice.

Questions

  • Which wardrobe tasks need instant offline feedback?
  • What data should never need to leave the device?
  • How much model quality is worth trading for lower latency?
  • How should model updates work across iOS and Android?

I will update this note as experiments produce results worth publishing.