Keyword-based job scoring falls in love with the wrong roles too easily

Situation

A job description can mention PHP, AWS, Python and React while the company is actually hiring an MLOps specialist, a junior support engineer or an architect. A bag-of-keywords score celebrates those overlaps even when the candidate's role experience does not match the work. False positives are especially expensive here because they consume limited human attention.

Approach

I split ranking into role-aware signals: title and seniority, core requirements, explicit technology evidence, responsibilities and negative specialization signals received separate weights. Junior overqualification and MLOps-style specialization gained regression cases. The goal was not to hand-code a rule for every edge case, but to make the structure of the score represent the role itself rather than just its vocabulary.

Outcome

The Best Match list became shorter but more relevant. Postings moved down when their technology text looked familiar but their day-to-day work did not. Ranking explanations also made it clearer why a role earned positive or negative weight.