How Girard Media runs a faceless explainer engine on self-hosted models

Challenge

Faceless YouTube and client explainers usually rent Midjourney + ElevenLabs + CapCut. That leaks margin and locks the cut into a flat file. Girard Media needed a commercial-clean, self-hosted pipeline that still opens in the CoreReflex editor.

Workflow

1. Write a retention script (script.json — hook → escalation → motive → payoff).
2. Kokoro VO per scene; SDXL flat-vector stills; ACE-Step underscore.
3. Assemble Ken Burns + kinetic captions on the GPU node; duck music under VO.
4. build_project.mjs emits an UndoableState project — VO, music, scenes, captions as separate stems.

Result

• No paid generative API on the production path when the GPU node is up.
• Licenses are Apache/MIT-class for the engines named above (not non-commercial audio models).
• The same assets can render a flat MP4 and open as an editable studio project.

Evidence

Operator-owned Epic Motive skill (.claude/skills/epic-motive/) and GPU fleet runbook. Full pixel proof requires a provisioned GPU node (docs/dev/gpu-inference-fleet-ops.md); unit tests cover package/edit contracts offline.