Ihza MahendraResume
Back to Production ML at scale
Production ML at scale

Pre-filter ranking layer above the main ranker

Ranking layer that pre-screens a large product pool so the strongest matches reach the main model first.

What it is

An additional ranking layer that sits above the main recommendation model, pre-filtering a much wider pool of candidate products.

What it's for

The main ranker can only score so many candidates inside its latency budget. A pre-filter layer expands the candidate pool the platform considers without breaking that budget, so the strongest matches actually reach the model that decides what to show.

How it was built

A lighter-weight ranking module feeding the final ranker, tuned to business goals so the strongest matches bubble up first. Drove +13% GMV, +9.24% orders, and +0.73% CTR.

My role

Co-built as part of the Algorithm team at ByteDance.

Built with
Deep learningRanking systemsDistributed inferenceByteDance cloudPythonC++

Want the full technical depth, the tradeoffs, what broke, what I'd do differently? Ask the agent about this project.