The agent team
Built as a flow, not a collection of tools.
OrchestratorQuinCEO
Work the Grader sends back returns to the Creative Director. It runs when it is needed, not on every campaign.
The model stack
Every job gets the right intelligence.
Quin mixes and swaps AI models per job — here’s what she runs on today, and roughly what a campaign costs to produce.
Primary intelligence
DeepSeek v4-flash — opens www.deepseek.com in a new tab
Writing, thinking, strategy, ad copy, reporting, and self-grading.
Knowledge retrieval
OpenAI Text Embedding 3 Small — opens platform.openai.com in a new tab
Document search and retrieval (RAG) across the brand’s knowledge base.
Image generation · via Replicate
Nano Banana 2 — opens replicate.com in a new tab
2K image generation — carries nearly all of the AI cost per campaign.
Visual quality assurance
Dola Seed 2.1 Turbo — opens www.volcengine.com in a new tab
Image QA and brand-check scoring before an ad ships.
Built to evolve
Modular by design.
Accountable by default.
Each job’s model is a setting, not a rebuild — we review the lineup roughly every 3 months and swap in a better model whenever one earns it.
Infrastructure