Skip to content

Demo Video Script

3-minute shot list for the Devpost submission video, recorded against the live deployed app:

https://premiere-control-room-web-sh76ssrjya-uc.a.run.app/

This deployment is confirmed running the real crew — GET /health on its backend returns agent_mode: "live", meaning real Gemini calls via Vertex AI and real Grafana Cloud MCP tool calls, not the deterministic mock crew. Say so on camera — it's the strongest claim this project can make, and it's true on this URL.

Known issue: verify Shot 4 works before recording

As of this doc's last check, injected incidents on this specific deployment were getting stuck at investigating and never progressing (the Detective step was failing silently — see Orchestrator._run_crew's swallowed exception handler in backend/app/orchestrator.py). Confirm this is resolved (check gcloud run services logs read premiere-control-room-backend --region <region> | grep -A 30 "Crew run failed" for a traceback if not) and that the Grafana dashboard is provisioned (infra/scripts/provision-grafana-dashboard.sh — see infra/scripts/README.md) before relying on Shot 4/5 working live on camera.

Before recording

  1. Sign in once, off-camera, and confirm your admin credentials work. Actions (inject anomaly, approve/reject) require an operator or admin account — see Setup Guide if you're locked out. Don't type or read the password on camera.
  2. Reset to a clean slate if you can. A fresh deployment has zero incidents (GET /api/analytics/summary → total_incidents: 0) — ideal. If the deployment already has incidents on it, that's not fatal (real data is arguably a stronger look than a scripted empty state), but re-check Shot 3's framing ("all healthy, nothing running") against whatever's actually on screen.
  3. This is a real system, not a fixture — budget real wall-clock time. Each agent turn is a real Gemini call plus real Grafana MCP tool calls; expect 5–25 seconds of visible "thinking" per step, not instant transitions. Two ways to handle this within a hard 3:00 cap:
  4. Record once, edit with speed ramps / jump cuts over the waiting periods (recommended — keeps it honest, still fits the runtime).
  5. Record the live run first, unscripted, to see real timing, then re-record narration to match the actual cut. Don't pre-script exact agent output text (root-cause hypothesis, brief wording) since the real crew's Gemini output varies run to run — the shots below cue off UI state changes, not literal on-screen text.
  6. Open the app in a clean, logged-out browser window at ~1440×1000 for the cold open, then sign in on camera per Shot 3.
  7. Have a second tab ready on the live Grafana Cloud stack (or skip Shot 5's deep-link if you'd rather not expose your Grafana org on camera — the embedded panel image alone still makes the point).

Shot 1 — Cold open (0:00–0:15)

Visual: Black slate, then a title card: "Premiere Control Room — an agentic reliability engineer for live media premieres."

Narration (≈13s spoken, rest is visual beat):

"A global streaming premiere breaks — CDN overload, encoder saturation, a bad origin deploy. Today, an engineer stares at a wall of dashboards while millions of viewers buffer. This is Premiere Control Room."

Shot 2 — Five agents, fast (0:15–0:35)

Visual: The architecture diagram (see docs/architecture.md) or a simple five-name graphic (Sentinel → Detective → Producer → Responder → Wrap). Highlight each name as it's said.

Narration:

"Five Google ADK agents, sharing one live Grafana Cloud MCP connection. Sentinel watches service-level objectives. Detective correlates metrics, logs, and traces into a root cause. Producer briefs the team in plain language. Responder proposes a fix. And Wrap writes the postmortem — automatically."

Shot 3 — Sign in, live control room (0:35–0:50)

Visual: The real deployed app. Empty Control Room, all agent tiles idle, Live QoE map all green ("healthy"). Click Sign in in the sidebar, sign in as operator/admin (credentials off-screen).

Narration:

"This isn't a mockup — it's live right now, running the real Gemini crew against a real Grafana Cloud stack. Everything's healthy. Let's break something."

Shot 4 — The core loop, live (0:50–1:50)

Visual: Back on the Control Room page, signed in. Click Inject demo anomaly.

  1. (0:50) Click the button; narrate as the Agent activity feed starts filling and the Sentinel/Detective/Producer status tiles light up blue ("running") one after another.

    "Sentinel flags a real SLO breach. Detective pulls real metrics, logs, and traces through Grafana's MCP tools to build a root-cause hypothesis. Producer turns that into an incident brief."

  2. (Whenever it lands — real Gemini latency, don't force a timestamp) the Responder tile turns amber, reading blocked, and the approval modal appears: "1 of 1 pending", the proposed action type, and Approve/Reject buttons.

    "And here's the moment that matters. Responder wants to act — but it's not allowed to. This isn't a prompt asking the model to be careful. It's a real function call, blocked in code, waiting on a human. The model can't talk its way past this."

  3. Click Approve on camera.

    "Approved."

  4. Cut to the incident's status flipping through to postmortem ready, and the Incident timeline panel populating with the full agent-by-agent history.

    "And Wrap closes it out — a full timeline and a written postmortem, generated the moment it resolves."

Shot 5 — Proof it's real (1:50–2:10)

Visual: Scroll to the Grafana panel card on the incident — a live-rendered image pulled straight from the real Grafana dashboard, with an "Open in Grafana →" link. Click it to briefly show the real Grafana Cloud UI in a new tab (optional — skip if you'd rather not expose your Grafana org).

Narration:

"That panel isn't a screenshot we baked in — it's rendered live from our own Grafana Cloud stack, the same one the agents just queried. This whole run — every model call, every tool call — is real."

Shot 6 — Concurrency and playbooks (2:10–2:40)

Visual: Back on the Control Room home. Click Inject 3 concurrent anomalies.

  1. Narrate over the feed as three incidents run in parallel and the Responder tile reads blocked ×2 (or similar — however many land as high-risk simultaneously).

    "Three incidents, three different playbooks. One's low-risk — the crew scales capacity and moves on by itself. Two are high-risk — CDN failover, a cache purge — and both wait for a human, tracked independently, one modal, one decision at a time."

  2. Approve one, reject the other, on camera — quick cuts, no narration needed over the clicks themselves.
  3. Let them settle to postmortem ready (or skipped for the rejected one).

Shot 7 — History & analytics (2:40–2:55)

Visual: Click History & Analytics in the sidebar. Show the stat tiles: Total incidents, MTTR, Resolved, Active, and — distinctively — Gemini input/output tokens and Estimated cost.

Narration:

"Every incident's archived automatically — mean time to resolution, breach frequency, and real Gemini token usage and cost, right down to the run we just watched."

Shot 8 — Close (2:55–3:00)

Visual: Title card: "Google ADK · Gemini · Grafana Cloud MCP · FastAPI · Next.js" plus the live URL and repo link.

Narration:

"Premiere Control Room. Reliability engineering that never blinks, and never acts alone."

Notes for whoever's cutting this

  • Total run time targets 3:00 flat. Shot 4 (the forced-approval gate) and Shot 6 (concurrent playbooks) are this project's two most differentiated claims — protect those if anything needs trimming, cut Shot 5 first if you're tight.
  • Because this is a live system, exact wording of the Detective's hypothesis, the Producer's brief, and the Responder's proposed action will vary run to run — don't script narration around literal on-screen text; cue lines off UI state (tile color, modal appearing, badge count) as written above.
  • Re-verify screen copy against the running app before your final cut — UI strings referenced here (Inject demo anomaly, blocked, 1 of N pending, postmortem ready, the Live QoE map's healthy/degraded/recovering labels, the History page's stat tile labels) were pulled directly from the frontend source as of this repo's current commit.