RouterCore: Safe Routing for AI Agents

RouterCore is a safe routing layer for agentic DevOps workflows. Before an AI agent or orchestrator executes a task, RouterCore decides whether a natural-language request should be routed, clarified, confirmed, rejected, or escalated. The system converts messy user requests into structured workflow routes and parameters, validates them against workflow schemas, applies an authoritative policy layer, and generates execution previews only. It supports workflows such as creating web apps, storage buckets, service accounts, scheduler jobs, and IAM role grants. For the AMD Developer Hackathon, RouterCore is framed as a Track 2 fine-tuning project. We trained a compact Qwen LoRA adapter on AMD Developer Cloud using ROCm PyTorch on an AMD Instinct MI300X. The safety-tuned model improved required-field extraction from 28.57% to 100.00% and status accuracy from 57.33% to 86.67%, while preserving 100.00% unsafe request rejection and 0.00% false route rate. RouterCore demonstrates safe routing, not just routing: fine-tuning improves the model’s structured proposals, while schema validation and policy checks remain authoritative before any agent handoff.