Let the small
model go first.
Oyster recommends skincare products. This classifier assigns categories such as cleansers and moisturizers. A wrong label can remove a useful product from the recommendation search.
- MY WORK
- End to end: model experiments, evaluation, the research fallback, and Laravel integration.
- WHAT CHANGED
- A small model handles clear predictions. Uncertain products get bounded research or remain deferred.
Decide when the model can act.
A category prediction and a catalog update are different steps. A score threshold decides which predictions can proceed automatically.
Where would you draw the line?
Select a record to inspect its decision.
The line controls which rows act, not how accurate the model becomes. These scores illustrate the rule; they are not measured predictions.
I compared text-embedding approaches with a classifier based on word and character patterns. Logistic regression provided a first pass without running a neural embedding model. Lower-scoring predictions could go to web research.
Accept predictions at or above the threshold. Leave the rest unresolved for the next step.
Even “defer”
has side effects.
A classification attempt should not make an unchanged product look newly edited.
I kept deferred attempts from changing the ordinary update timestamp, avoiding an unnecessary trip through re-embedding.
See how Oyster uses those categories