Small Language Models Demonstrate On-Prem HR Triage

This article demonstrates how to implement privacy-first "Local First, Cloud Last" agents using small language models (1–3B) for an on-premises HR triage system. It details a multi-model pipeline—MiniLM embeddings for intent detection, Phi-3-mini for planning, and Function Gemma for constrained function execution—running on standard hardware and executing end-to-end within roughly 10–30 seconds. The repo, file descriptions, and execution logs illustrate practical deployment steps for enterprises with strict data-locality requirements.