# Joe — NVIDIA Agent Challenge Demo Script (91s)

**Format:** 1920×1080 · H.264 · yuv420p · 30fps · silent (subtitles only)
**Theme:** Joe: Nemotron-Powered Edge AI Agent Factory
**Style:** NVIDIA tech-demo dark theme · neon-green #76b900 · cyan #00ffff · purple #7b61ff
**Repo:** https://github.com/applelai001/joe

| # | Time      | Card               | Narration (EN)                                                                                  | 中文                                              |
|---|-----------|--------------------|-------------------------------------------------------------------------------------------------|--------------------------------------------------|
| 1 | 00:00–00:13 | Cold Open          | Joe. A Nemotron-powered Edge AI Agent Factory. 100+ models, 100+ skills, 6 nodes, one brain.   | Joe — 一個 Nemotron 驅動的地端 AI Agent 工班。   |
| 2 | 00:13–00:26 | The Problem        | Enterprise AI fails on three things: vague requests, heterogeneous GPU fleets, and zero verification gates. | 企業 AI 三大痛點：需求模糊、硬體分散、缺乏驗證關。|
| 3 | 00:26–00:39 | Architecture       | Joe dispatches across RTX 5090, Arc A770, and an x86 fleet, with Copilot as the verifier.       | Joe 在 RTX 5090、Arc A770 與 x86 多節點派工，Copilot 把關。|
| 4 | 00:39–00:52 | Live Capabilities  | Real portal, live `/healthz`, five Copilot terminals — all WAN-aware behind one nginx.          | 真實前台、`/healthz` 即時健康、五個 Copilot 終端、WAN 自動路徑。|
| 5 | 00:52–01:05 | The Loop           | Plan, dispatch, execute, verify, learn — every task ends with a verifiable delivery package.    | 計畫→派工→執行→驗證→課程，每次都交可驗證包裹。  |
| 6 | 01:05–01:18 | Verifiable Delivery| PPTX deck, capability registry, WAN smoke log, ffprobe report, Iron Law 40.5 proof, lesson.    | 簡報、技能清單、WAN 巡檢日誌、ffprobe 驗證、課程紀錄。|
| 7 | 01:18–01:31 | CTA                | Joe is open source. Built for NVIDIA Agent Challenge by joyce628.                                | Joe 開源。為 NVIDIA Agent Challenge 而生。       |

## Real evidence shown
- Card 4 embeds real screenshots: portal (login wall) + `/healthz` JSON live response
- Card 3 topology reflects actual SSH-registry fleet (1.98 RTX 5090 · 1.80 Arc A770 · 1.63 · 1.118 · 1.210 · 1.10)
- Card 6 paths all exist on disk

## Pipeline used
Pillow → 7 PNG cards (1920×1080) → ffmpeg loop-encode each card to 13s clip @ libx264 crf 20 → concat demuxer → mp4 yuv420p.

## Verification (Iron Law #40.5)
```
ffprobe → duration=91.000  codec_name=h264  width=1920  height=1080
size    → 1.2 MB (static-card slideshow, highly compressible)
frames  → 0:15 / 0:60 / 1:30 sampled OK
files   → 4 / 4 deliverables present
```