Can AI Understand Capability Beyond Keywords? An Agentic RAG Architecture with Privacy Controls for NeuroJobs.ai and NeuroRecruiter.ai

This work presents an agentic retrieval-augmented generation (RAG) architecture with privacy controls for NeuroJobs.ai and NeuroRecruiter.ai. The architecture treats hiring as a constraint-aware evidence retrieval and workflow problem rather than a simple keyword-matching task. The proposed system introduces Identity and Capability Separation (ICS) to separate direct personally identifiable information from capability data before semantic indexing. It combines lexical retrieval, dense vector search, metadata filtering, deterministic constraints, evidence-grounded generation, specialized agents, and human approval gates. The reference architecture maps orchestration to Amazon Bedrock AgentCore, exposes approved tools through Model Context Protocol (MCP), uses Qdrant for vector retrieval, and separates ephemeral cache from durable workflow checkpoints. The work also reports scoped engineering verification for NeuroJobs.ai, including Lighthouse accessibility scores of 100/100 on the tested Homepage and Login pages and a median mobile performance score of 98. These measurements validate the delivery surface and are not presented as evidence of candidate-ranking accuracy, hiring outcomes, recruiter productivity, or fairness. A four-system evaluation framework compares BM25, Dense Retrieval, Hybrid RAG, and NeuroMatch Agentic RAG, covering retrieval quality, grounded generation, recruiter workflow efficiency, privacy, fairness, adversarial robustness, latency, and cost. The proposed architecture is designed as a decision-support system rather than an autonomous hiring system, with consequential employment decisions remaining under human authority.Project platforms:NeuroJobs.ai — Neurojobs.aiNeuroRecruiter.ai — Neurorecruiter.aiNeuroWhale — Neurowhale.ai