Job Ad Classification at Scale

Job adverts are written by humans, so the same role appears as "Senior Dev", "Software Engineer II" and "Full-Stack Ninja". To make labour-market data comparable, every advert has to map to one standardised occupation code.

Doing that at six hundred million adverts is where it gets hard. The obvious approach — compare each advert against every occupation — was both too inaccurate to use and too expensive to run.

A two-stage ranking pipeline. A fast bi-encoder narrows millions of possibilities down to a shortlist, then a slower, more careful cross-encoder re-ranks that shortlist to pick the winner. This is the standard retrieval-then-rerank pattern, and it is what makes the accuracy affordable.

The accuracy gain came mostly from hard-negative mining: training the model on more than 2.5 million pairs of examples that look similar but are genuinely different occupations. Easy examples teach a model very little.

Cost came down through FP16 inference and partitioning the workload so autoscaling GPU workers stayed busy instead of idling.

Top-1 accuracy went from 12.48% to 65.13%; end-to-end pipeline accuracy from 10.47% to 62.24%.

The cost of a full 600M run fell from a projected $48,000 to roughly $12,000, and runtime from about 23 days to about 4.7 days.

Validated in production on 20.9 million records in around 12 hours, then run across the full corpus of 600 million adverts.