Meridian: A Unified Multi-Agent HR System for Recruitment, Onboarding, and Performance Review

Many modern HR software tools tend to use fragmented approaches, with three separate solutions for recruitment, onboarding, and performance management, each with its own databases and lacking end-to-end candidate/employee data. We present Meridian, a proof-of-concept multi-agent system for managing recruitment, onboarding, and performance reviews of employees as an end-to-end process under a unitary relational model, in which candidate records are passed through LangGraph-based agents performing resume parsing, hybrid candidate scoring, email/welcome-kit personalization, training-plan synthesis, peer feedback summation, and draft performance review composition, with additional retrieval-augmented policy chatbot and workflow orchestration layers for day-to-day operations. Our open-source prototype relies on free tiers and local tools whenever possible, including Groq-hosted Llama 3.3 70B or Gemini for text generation, Sentence Transformers and ChromaDB for retrieval, SQLite as a relational database, and n8n for workflow management, thus avoiding the “enterprise pricing” lock that would make our prototype inaccessible to smaller companies experimenting with agent-assisted recruitment pipelines. In this paper, we describe the technical design of our prototype, focusing on the explainable hybrid candidate scoring algorithm, which combines keyword-based matching, semantic comparison with the target job description, and years of experience in a weighted manner, and the human-in-the-loop design of our agents’ operations. Furthermore, we demonstrate how Meridian could be used in practice, applying one of our agents to transform raw candidate resumes into a ParsedResume object containing structured data for downstream processing, and discuss our plans for further development and evaluation, which include a limited pilot study with actual candidates, transformation of parsed resumes into onboarding records with automated training-plan synthesis, and an analysis of candidate scores with a focus on the relative impact of different fields. Lastly, we discuss the limitations of our approach as a whole and potential directions for future improvement.