Applied AI / Creative Production
Static Ad Production Studio
An evidence-bound creative production workspace that turns an approved direction into static-ad versions, pinned regional feedback, exact-version review and deterministic placement packages — with copy and safe areas preserved by check, not by hope.
An independently branded, sanitized reconstruction informed by engineering work performed as Founding Lead Engineer at Peerless CMO. Not Peerless production software.
The hosted frontend runs the complete deterministic synthetic workflow. No provider key, live ad account or publishing integration is connected.
01 — The problem
Generated creative gets the picture right and the words wrong.
Image models are good at composition and unreliable at text. Ship them into ad production and you get plausible-looking creative with mangled copy, logos in the safe area, and geometry that breaks the moment it hits a placement with different dimensions.
Review is the other half of the problem. Feedback arrives as "the headline feels cramped" attached to nothing in particular, against a version that has already been superseded twice.
02 — The system
Explicit inputs, deterministic checks, immutable versions.
Approved strategy, safe claims, brand rules and asset roles are explicit inputs rather than prompt context. Deterministic checks preserve exact copy, safe areas and placement geometry. Reviewer feedback is pinned to normalized regions and creates immutable child versions, and approved output is packaged idempotently for placements without ever publishing externally.
Explicit inputs
Strategy, safe claims, brand rules and asset roles are structured inputs, not prose in a prompt.
Exact-copy checks
Rendered text is verified character-for-character against the approved copy.
Safe-area & geometry checks
Placement geometry is validated deterministically, per placement.
Pinned region annotations
Feedback attaches to a normalized region of a specific version.
Immutable child versions
An annotation produces a new version; the reviewed one never mutates.
Idempotent packaging
Approved output packages once per version and never publishes to an ad account.
03 — Architecture
How a version reaches a placement package
04 — Engineering decisions
Choices that keep generated creative shippable.
Copy is verified, not trusted
Text rendered by an image model is checked character-for-character against the approved copy. This is the single check that separates a usable ad pipeline from an expensive random image generator.
Geometry is a deterministic test
Safe areas and placement dimensions are validated by code, not judged by eye in a review call. The failure mode they catch is the one that only shows up after the media buy.
Annotations are normalized regions
Feedback is pinned to coordinates that survive resizing, so "this bit" still means the same bit on a different placement.
Review targets an exact version
Annotations create immutable child versions. A reviewer never approves a moving target, and history is reconstructable.
Optimistic concurrency on creative and review
Stale mutations are rejected. Parallel review is normal; silently losing a reviewer’s comment is not.
Packaging never publishes
The system produces local packages. No path exists to push to a live ad account — the boundary is architectural, not procedural.
05 — Reliability & controls
What the system does when things go wrong.
06 — Evidence
Verified results
None published. No client creative performance is claimed on this page.
Benchmarked results
Public dataset: Deterministic evaluation set with full verification pipeline
Delivery tier: Tier 1 runnable repository
Verification: lint · web tests · build · ruff · pytest · evals
Supply chain: OpenAPI generation · secret scan · pnpm audit · pip-audit
The suite asserts production behaviour — exact copy preservation, geometry, concurrency, approval and tenancy — rather than aesthetic quality. It requires no Supabase, Railway, image-provider key or live ad account.
Simulated results
All included brands, assets, claims, creative results, comments, approvals, checksums and experiment records are synthetic.
Projected business value
Not claimed. No performance, ROAS or conversion figure is attached to this system.
Verified, benchmarked, simulated and projected figures are kept separate on purpose. A benchmarked number is never presented as a verified client result.
07 — Technology
Technology supports the story. The architecture and the controls are the story.
08 — Repository & demo
Generated creative that keeps failing review?
We can put deterministic copy, safe-area and geometry checks in front of review, so only shippable versions get there.