feat: add transcript for demo video detailing ARIA's functionality and use cases

Co-authored-by: Copilot <copilot@github.com>

Author: zestones

docs/demo/video.md

@@ -0,0 +1,54 @@
+# Transcript of the demo video
+
+## ACT I — The problem
+
+In every factory, every plant, every water station in the world — there's one person who knows.
+He knows when a machine sounds different. He knows it's going to break — two days before it does. He just knows. And when he retires, that knowledge disappears. Forever.
+
+Companies have tried to fix this for ten years. Setting it up costs half a million dollars and takes six months of specialists. So ninety-five percent of industrial sites just... don't. They wait for machines to break. We built ARIA to change that. Five agents, each with a single job passing the problem between them. Exctly like a real maintenance team passes a ticket.
+
+## ACT II — ARIA at work
+
+### Scene 1 — Onboarding Bottle Labeler
+
+Take a water-bottling factory. One line, four machines. Today a fifth one comes online — the Bottle Labeler. Normally, configuring its monitoring takes a specialist two days. Drop in the manual. ARIA reads it. Asks three questions. Two minutes later, it's live.
+
+### Scene 2 — Forecast a breakdown
+
+Meanwhile, on the Bottle Capper — ARIA flags a potential breach. But look at what happens next. Vibration is falling, not rising. ARIA reads the context, checks the knowledge base, and concludes: no action required.
+
+It didn't just detect — it judged.
+
+### Scene 3 — Investigating an Anomaly
+
+#### 3a · Sentinel breach + thinking
+
+Now let's see what happens when a real breach hits.
+
+The operator sees an alert — and that's where most systems stop: alert sent, problem yours.
+
+ARIA doesn't.
+
+It fires the Investigator agent — and like a detective, it starts gathering clues: operator logbook, sensor trends, equipment knowledge base.
+
+### 3b · Sandbox Python
+
+And here's the part that cannot happen without Managed Agents.
+
+The agent wrote Python, ran it live inside Anthropic's cloud sandbox, and computed the degradation rate directly from the raw signal data.
+
+Slope: zero point three two millimeters per second per hour. R-squared: one point zero zero zero. Time to trip threshold: four point five hours.
+
+That's not a language model guessing — that's a regression that actually ran.
+
+### 3c · RCA
+
+And those exact numbers land in the work order.
+
+Discharge bearing wear — progressing thirteen point seven times faster than the January 2026 baseline. Twelve steps. The exact part number.
+
+A four-hour maintenance window, starting midnight. The technician walks out with a printed sheet.
+
+### Scene 5 — But How?
+
+