Employee Referral System

The referral pipeline at Hexaware was entirely manual — HR reviewed hundreds of PDFs per week to decide whether a referred candidate fit a JD. The process was slow (avg 5 days to first feedback), inconsistent, and created a poor referrer experience that reduced future referral quality.

Built an ML-driven matching engine that parses uploaded résumés (PDF/DOCX), extracts skills and experience vectors, and scores them against structured JDs using cosine similarity + a gradient-boosted ranking model. Recruiters get a ranked shortlist with per-candidate fit explanations. Referrers get a match score immediately on upload. Replaced the manual first-pass entirely, cutting time-to-first-feedback from 5 days to under 2 hours.

• 01

  Résumé parsing quality is the highest-leverage problem — garbage in means garbage match scores regardless of model sophistication.

• 02

  Explainability matters as much as accuracy for recruiter adoption: 'this candidate scored 0.82' means nothing; 'missing: distributed systems, has: Python, ML' gets buy-in.

• 03

  Building an internal feedback loop (recruiter accept/reject signals) let the model improve over time without a separate labelling workflow.

• 04

  React dashboards for non-technical users need to ruthlessly hide complexity — early versions had too many knobs.