AI Engineer — Building a Creative Intelligence & Ad Generation Engine

How the system works We are building a programmatic creative intelligence and rendering engine for performance advertising. This is explicitly not an AI image-generation system. The engine does not ask an image model to generate a new advertisement from a prompt. Instead, brands provide their real creative assets — product photography, packshots, lifestyle imagery, logos, fonts and other approved brand materials. The system then combines several layers of information: Brand system Colors, typography, spacing, visual rules, tone of voice, constraints and other rules that define what the brand should look and sound like. Commerce data Products, variants, pricing, inventory, offers and other factual product information from systems such as Shopify. Creative references Existing brand creatives and selected advertisements from other brands. Computer vision analyses how these creatives are constructed: image combinations, text placement, hierarchy, grids, overlays, product cards, pricing layouts, negative space and other visual mechanisms. References are used to understand creative structure, not to copy another brand's visual identity. Performance data Historical advertising data can be connected back to individual creatives so the system can understand which hooks, concepts, layouts and creative mechanisms performed well. The intelligence layer uses this context to decide what should be created. For example, it might determine: Use this product and these two supplied product images. Build a multi-image product composition inspired by a structure observed in the reference set. Use a price-focused concept because the offer is currently active. Apply Delarue's typography, spacing and brand rules. Generate these copy elements from verified product facts. That decision is then passed to a code-based rendering engine. The renderer programmatically constructs the advertisement using the supplied assets. Images are cropped and positioned, text is typeset, modules are arranged, backgrounds and overlays are created, prices are formatted and layouts are adapted to different aspect ratios. The result is a normal rendered image, but its construction is deterministic and code-driven. Why this architecture matters We want to be able to generate hundreds or potentially thousands of creative variations from an approved set of brand assets. We do not want to make hundreds of manually designed templates, and we do not want to make thousands of expensive generative-image API calls. Instead, the system should work more like a creative grammar. Creative families define the types of advertisements that can exist. Modules define elements such as images, headlines, prices, badges, product cards and logos. Composition rules determine how those modules can be combined. Brand rules determine what is allowed. References provide examples of useful visual mechanisms. Performance data influences which combinations should be explored. This creates combinatorial variation: concept × layout × module configuration × asset selection × copy × product × format A relatively small set of reusable primitives can therefore produce a very large number of genuinely different creatives. Once the intelligence decision has been made, rendering another variation is essentially a software operation rather than another generative-AI job. This makes high-volume creative production fast, deterministic and extremely inexpensive at the rendering layer. AI is still important, but it is used primarily to understand, reason and make creative decisions — not to hallucinate the final pixels. The feedback loop The longer-term system closes the loop: Brand + assets + references + product data + historical performance → Creative intelligence → Concept and composition decisions → Programmatic rendering → QA and validation → Ads launched → Performance captured → System learns which mechanisms work → Next generation of creatives Every generated creative also retains its lineage: which product, assets, facts, reference mechanisms, creative decisions and rules produced it. That allows performance to eventually be attributed not only to "Ad 147", but to the underlying creative decisions that produced Ad 147.