PromoRail: from promo brief to brand-safe Google Ads execution
A governed workflow that combines prior-year memory, current-year facts, deterministic copy checks, Google Ads Editor output, and dry-run-first API control. Built by a paid-media operator to remove repetitive work without surrendering judgment or safety.
Problem
Every promotion, building the ads meant combining historical research, on-brand copywriting, bulk spreadsheet mapping, QA, and execution risk — hundreds of ads across a big product catalog, under time pressure. Done by hand it's slow and error-prone, and a single mistake fails the whole bulk import.
Insight
Prior work should become structured memory the system can reuse — but stale facts (last year's dates, discount, and landing paths) must never be copied blindly. Separate the reusable patterns from the volatile facts, and regenerate the facts every time.
System
An ingested promo calendar (promos preloaded with their dates) → a parsed marketing brief (facts extracted with provenance) → prior-year memory → a guideline-compliant copy bank → a deterministic validation gate → a realistic preview → controlled output → year-over-year measurement. AI can propose copy; the same rules that gate the seed also gate anything AI writes.
Safety
No direct trust in model output. Every asset is validated deterministically, the export is a Paused Google Ads Editor CSV, and the API path is an interactive, dry-run-first console — it scans, resolves the 3-enabled-RSA limit by pausing the lowest-performing ad, surfaces policy issues, and stages labeled paused ads for a human to review and enable. Nothing goes live automatically.
Outcome
A repeatable workflow that compresses hours of manual work into minutes, removes whole classes of bulk-import errors, and is teachable to another operator — the same institutional knowledge, encoded as software.
What it proves
Product thinking, marketing domain depth, AI governance, software delivery, and enablement — the operator-builder skill set, made inspectable rather than merely claimed.
What stays private
The public twin is fully synthetic. The production application, real advertiser data, live account structure, credentials, and exact production copy remain private. This demonstrates the workflow — the product judgment, domain depth, and governance — not the confidential system it mirrors.