SAGE: Fine-Tuned Small Language Models, Multi-Agent Orchestration, and Governance for Digital Marketing Automation

Abstract

Cloud-hosted Large Language Model (LLM) agents offer broad capability but limited control over per-token cost, external data exposure, and vendor dependency. This work addresses that trade-off by decoupling supervisory reasoning from domain-level inference. This work presents SLM Agentic Governance Engine (SAGE), a unified framework in which a cloud-hosted LLM performs workflow decomposition, agent routing, and response synthesis, while locally fine-tuned Small Language Models (SLMs) and predictive models perform all domain-specific classification and prediction on the deployment host. Governance controls using role-based access control (RBAC), JSON-line audit logging, and OpenTelemetry (OTel) tracing are embedded in the orchestration layer at design time rather than instrumented per agent. The framework is instantiated for multi-task digital marketing automation using Google Agent Development Kit (ADK), comprising five Low-Rank Adaptation (LoRA)-adapted Qwen 2.5-0.5B models, a fine-tuned DistilBERT sentiment classifier, and an Extreme Gradient Boosting (XGBoost) churn predictor coordinated by a Gemini-3.1-Flash-lite supervisory agent. Because domain inference is local, customer data never leave the organisation during task-level prediction, and each model can be retrained and redeployed independently of the orchestration layer. Evaluated on six public datasets, the system achieves task accuracies of 74.0%–98.0%, sentiment macro receiver operating characteristic area under the curve (ROC-AUC) of 0.932, churn ROC-AUC of 0.846, and a mean end-to-end latency of 3.9 s, on commodity hardware without a discrete GPU.