VERDICT: Evidence-First Candidate Evaluation

Capable candidates get rejected by broken evaluation systems that reward presentation over substance. No feedback. No transparency. No accountability. Just silence. VERDICT was built to fix that. VERDICT is a multi-agent candidate evaluation system where four specialized AI agents collaborate through Band to assess candidates across research program admissions, corporate hiring, and compliance screening. Each agent processes the candidate profile in sequence and posts its output to a Band room, handing off context to the next agent in the pipeline: - Evidence Extractor parses the candidate profile for verifiable, concrete signals — projects, skills, experience, achievements. - Criteria Mapper scores that evidence against explicit role requirements with per-criterion justification and an overall fit score. - Bias Auditor flags irrelevant evaluation factors — geographic bias, institutional prestige bias, GPA cutoff rigidity, publication gatekeeping. - Accountability Agent produces a mandatory structured verdict (ADVANCE / WAITLIST / REJECT) with full evidence citations, candidate feedback, and an improvement roadmap. The system was validated with a real candidate profile: a Biomedical Engineering undergraduate applying to the MITACS Globalink Research Internship. VERDICT returned a WAITLIST verdict at 80% confidence, caught three bias flags, and produced a 70/100 signal score — all with full reasoning visible in the Band feed. Built on Flask, LangGraph, LangChain, Groq (Llama 3.3 70B), and the Band REST API. This system exists because the current process failed someone who deserved better. VERDICT cannot give back what was lost. But it can make sure the next candidate gets an answer they can actually use.