A pragmatic recommendation system built on the client's existing warehouse, with an evaluation harness the in-house team now owns.
Merchandising was manual and generic. Previous ML attempts stalled because nobody could tell if they were working.
We defined success metrics first, built an offline evaluation harness, then shipped a simple co-visitation model before layering in learned ranking.
The pilot cohort saw an 18% revenue lift. The team runs weekly evaluations and ships model changes themselves.
The co-visitation baseline captured most of the value and set a bar every later model had to beat in evaluation before shipping.
Have a system like this?
Start a conversation