agents
agents
Agentic HR is a LangGraph-orchestrated HR assistant designed for environments where answers must be grounded and actions must be controlled. It combines policy-aware RAG, tool-calling into HR systems, and a mandatory human approval path for risky operations.
In short: this is not a generic chatbot. It is an execution framework for HR workflows where correctness, safety, and auditability matter as much as language quality.
Pulls grounded context from HR policy PDFs so answers stay anchored to company rules, not model guesswork.
Routes action-oriented requests to system tools for records, collaboration, and task operations.
Pauses sensitive or destructive actions and requires explicit human approval before execution.
The assistant starts with retrieval over embedded policy documents before drafting a response. This sharply reduces hallucinations in compliance-sensitive HR scenarios.
Beyond Q&A, the graph can trigger structured tool calls against HRIS-style systems and internal workflows. It can read, reason, and act within one conversation loop.
Requests and model outputs run through safety checks for PII leakage, prompt-injection patterns, and policy violations before anything reaches end users.
Potentially destructive actions are intercepted and parked in an approval queue. Human reviewers can approve or reject with reason codes through API endpoints.
Conversation state is checkpointed in Postgres by session id, so context survives restarts and users can continue from where they left off.
The architecture is intentionally auditable: traceable decisions, explicit tools, deterministic approvals, and a clear boundary between retrieval and action.