demo-script.md

2-minute demo script (live or recording)

Goal: Show trust, grounding, and escalation — not feature laundry.

Setup (5 seconds)

• Repo with data/ present; pip install -r code/requirements.txt.
• export ORCHESTRATE_DISABLE_LLM=1 (repeatable) or show both offline + LLM if quota allows.

Minute 0:00–0:30 — Happy path (grounded reply)

1. Point to support_tickets/sample_support_tickets.csv — realistic ticket.
2. Run from repo root: python code/main.py --limit 1 (or scripts/run_agent.sh) writing to a temp output.
3. Open row: show status=replied, justification citing retrieval path/score language.

Say: “Answers are composed from retrieved chunks only in offline mode.”

Minute 0:30–1:00 — Escalation (risk)

1. Show risk.py pattern example (grading dispute / outage wording) or a ticket that triggers escalated.
2. Show status=escalated and generic safe response — no invented policy.

Say: “We escalate before we invent.”

Minute 1:00–1:30 — Invalid / out-of-scope

1. Short gratitude or trivia-style row → request_type=invalid or out-of-scope reply.

Say: “Invalid bucket avoids wasting human escalation.”

Minute 1:30–2:00 — Engineering credibility

1. python -m pytest code/tests -q
2. python code/run_eval.py --offline — routing metrics on sample.

Say: “Regression + CLI resilience so iteration doesn’t rot quality.”

One sentence close

“We optimized trustworthy routing and corpus-grounded text, not parametric knowledge—and we know exactly where paraphrase and multilingual inputs degrade.”