The LLM never saw the whole job market — only the uncertain cases left after deterministic filtering

Situation

Sending every job posting through an LLM would have been an obvious design, but it would also have been expensive, slow and difficult to reproduce. Many decisions did not need an LLM at all: an explicitly onsite US role or a completely different profession can be handled deterministically. AI was useful where the requirement language itself was genuinely ambiguous.

Approach

Role-aware deterministic ranking and eligibility ran first. Only a bounded top or uncertain set went to grounded AI review, where the model had to structure must-have evidence from the actual job text in a traceable way. The output was schema-driven, and the system preserved the deterministic base score and the AI review as separate sources rather than collapsing them into one opaque result.

Outcome

AI added value where responsibilities and true mandatory requirements were hard to separate linguistically, but it never became the pipeline's sole decision-maker. The overall system remained faster, cheaper and easier to diagnose.