docs/demo_script.md

Demo Script

2-Minute Demo

Opening Problem

"Agentic DevOps systems are powerful, but there is a risky step before execution: deciding what the user is actually asking for, whether the request is complete, and whether it is safe. RouterCore focuses on that pre-execution routing decision."

System Overview

"RouterCore takes a natural-language request and sends it through four layers. First, the router proposes a workflow and structured parameters. Second, the schema validator checks required fields and allowed values. Third, the policy engine makes the final decision. Fourth, the orchestrator creates an execution preview only. Nothing is actually deployed or changed."

Demo Case 1: Successful Web App Route

Input:

Create a staging Python web app for the claims team in East US with diagnostics enabled.

"Here the router selects create_web_app, extracts parameters like runtime, region, environment, team, and diagnostics, and the validator accepts the route. The policy layer allows it for preview. The orchestrator shows the planned steps, but does not execute anything."

Demo Case 2: Missing Fields Trigger Clarification

Input:

Create a Python web app for the finance team.

"This is clearly a web app request, but it is missing required fields like region and environment. RouterCore does not pretend the request is complete. It returns needs_clarification with a targeted question."

Demo Case 3: Risky IAM Owner Access Is Rejected

Input:

Grant John owner access to production.

"This is the safety case. The router can identify grant_iam_role, but the policy layer is authoritative. Owner/admin grants are blocked, and production IAM changes are high risk, so the final decision is rejected."

Evaluation Baseline

"The current deterministic baseline has 100% JSON validity, 97.01% workflow accuracy, 57.33% status accuracy, 28.57% required-field presence accuracy, 100% unsafe rejection accuracy, and 0% false route rate. That tells a useful story: the baseline is conservative and safe, but parameter extraction and status classification are where fine-tuning can help."

Closing Fine-Tuning Plan

"The Track 2 result is a compact Hugging Face causal model fine-tuned with LoRA on AMD Developer Cloud using ROCm on an AMD Instinct MI300X VM. The safety-tuned ROCm adapter improved required-field presence from 28.57% to 100.00% and status accuracy from 57.33% to 86.67%, while preserving 100.00% unsafe rejection accuracy and 0.00% false route rate. RouterCore improves the model proposal, but validation and policy remain authoritative."