AI Candidate Screening — Signal-Based Evaluation at Scale

We historically evaluated candidates using 1–5 Likert ratings on Airtable. The problem: reviewers gravitated toward average scores to avoid making difficult calls, or because the evidence wasn’t clear-cut. A candidate rated 3/5 on “technical depth” tells you almost nothing—it’s a hedge, not a signal.

The scores were also hard to compare across roles and organizations. What counts as a 4 for one reviewer is a 3 for another. Re-screening when requirements changed meant re-reading every profile and re-calibrating the same ambiguous scale.

Replaced Likert scales with binary signal evaluation. Each candidate is scored against specific, testable variables—has_shipped_python_in_production, has_published_at_top_20_conference, has_completed_cs_phd—with three possible verdicts: pass, fail, or unknown. Claude evaluates each signal with evidence citations drawn from CVs, LinkedIn, GitHub, and personal websites.

Binary signals are definitive and actionable—no room for fence-sitting. General variables like has_completed_cs_phd work across roles and orgs, while role-specific ones can be added as needed. The result: a recruiter can scan a candidate’s signal profile and immediately see who to shortlist, without re-reading source material or second-guessing ambiguous ratings.