Evidence-Grounded Resume Screening Copilot

Delivery approach

1. S

   Situation

   Recruiters screened large resume batches with inconsistent formats, titles, skill names, and levels of detail. Keyword filters missed equivalent experience and rewarded repeated terms. A single unexplained score would give recruiters no way to see why a candidate matched or correct a poor interpretation.

2. T

   Task

   Build a recruiter-facing assistant that processed more than 1,000 resumes per month and exceeded 90% screening accuracy against agreed role-fit criteria. Recommendations needed source evidence and configurable job requirements. Recruiters would control the shortlist, and model output alone could never reject a candidate.

3. A

   Action

   Each resume is parsed into employment, education, project, certification, and skill records while page and text offsets are retained as evidence. Skills and titles are mapped to a controlled vocabulary without discarding the original wording. LlamaIndex builds the candidate context, and Pinecone matches role requirements against demonstrated experience. Scoring keeps must-have requirements, preferred evidence, recency, and depth separate instead of reducing them to keyword frequency. Recruiters can change the criteria and open the passages behind each signal. Calibration uses recruiter-labeled examples; disagreements lead to a review of the definition before any score tuning. Test resumes include unusual layouts, missing dates, career gaps, transferable skills, prompt-like text, and near-duplicates. Production monitoring covers parse failures, evidence coverage, score drift, overrides, and search latency. Role-based access and retention rules apply to candidate data.

4. R

   Result

   The system processes more than 1,000 resumes per month with over 90% accuracy on the defined evaluation set. Recruiters receive comparable summaries and can open the evidence behind a match before adding someone to a shortlist. Override tracking shows where role criteria or extraction rules need correction, while the final decision remains with the recruiter.