MatchScore — LLM Agent for JD ↔ CV Matching (Custom GPT)

Live Agent: https://chatgpt.com/g/g-6973b0c982ec8191a540f05cd1d20737-matchscore

Overview

MatchScore is an LLM-based agent that compares a Job Description to a CV and produces:

• a structured match score
• strengths & gaps
• actionable recommendations (what to improve / highlight)

This repo documents the agent specification, guardrails, and an evaluation framework.

Why this project

The focus is reliability + safety + evaluation for AI agents:

• structured outputs (consistent format)
• reduced hallucinations (evidence-first approach, where applicable)
• bias-aware behavior (avoid protected attributes, where applicable)
• scenario-based testing and failure-mode analysis

How to use (high level)

1. Provide a Job Description
2. Provide a CV (use anonymized/synthetic CVs only)
3. Review the structured output: score + gaps + recommendations

Repo Structure

• agent_spec/ — agent instructions, scoring rubric, output schema
• eval/ — test cases + results template
• examples/ — anonymized sample inputs/outputs
• docs/ — threat model and limitations

Evaluation (what we test)

See eval/test_cases.md. Key dimensions:

• Consistency: same input → stable score/output
• Must-have vs Nice-to-have: correct weighting
• Hallucination control: claims supported by input (where applicable)
• Safety & fairness: avoid sensitive/protected attributes
• Prompt injection resistance: ignore malicious instructions embedded in inputs

Privacy / Data Handling

Do not upload real personal data (real CVs, emails, phone numbers, IDs). Use anonymized or synthetic examples only.

Notes / Limitations

See docs/limitations.md for known limitations and future improvements.