Automation system · 2026
UGC Ad Factory V3
A tested UGC production pipeline that keeps creative reasoning flexible while moving approval, spend control and delivery validation into deterministic boundaries.
- Discipline
- AI automation & media operations
- Built with
- n8n · ElevenLabs · OpenAI · Gemini · Firecrawl
01 / Why rebuild it
The problem
The first version had grown into roughly 110 workflow nodes. It could produce media, but too much orchestration lived in one system and an AI agent sat too close to the point where real provider credits could be spent. The rebuild focused on making state, approvals and money movement easy to reason about.
02 / System design
Four responsibilities
V3 separates creative work from generation and delivery so each boundary can fail safely.
- Product evidence → creative generation → independent judge → human approval
- Exact provider configuration → estimate → code-enforced budget gate → one generation
- Deterministic completion polling and output-contract validation
- Optional market intelligence for positioning, never copied competitor creative
03 / Spend boundary
Code decides whether money can move
The provider node is configured first with the requested model, duration, aspect ratio, resolution, audio setting and approved references. That exact node is estimated. Code then checks the returned cost and provider blockers. Only the true branch of that deterministic gate can reach the paid generation call. No language model decides whether spend is allowed.
04 / Tested result
Failure first, then one paid run
A zero-budget regression returned an estimate of 2,726.7273 credits and executed the paid generation node zero times. A second smoke test with a 3,000-credit ceiling submitted exactly one generation at the same estimate, approximately $0.50 at the provider's reported pricing.
- Seedance 2 Fast
- 9:16 portrait
- 480p
- Audio enabled
- Requested 4 seconds; delivered 4.096 seconds
- Output contract passed before final QA
05 / Delivery rule
Completed does not mean deliverable
Provider completion is only the start of delivery validation. The finished asset is checked against the requested model, aspect ratio, resolution, duration and audio contract. Only a valid result can move to final human QA. Regeneration never starts automatically from the QA stage.
06 / Current status
Core proven; client layer still being finished
The creative, approval, deterministic generation, polling and final-QA boundaries have been tested. The remaining work is deliberately boring: secure the client-facing API, connect the request interface and activate the production workflows. I am not presenting this as a finished SaaS until that layer is complete.
System views
The actual workflows.
Automation work