What I built
Marketing Execution is a local-first desktop product for performance marketing teams. It turns a meeting transcript into a reviewable campaign-change proposal instead of treating a conversation as permission to spend money.
A user imports an English or Korean transcript. The product extracts campaign decisions, action items, hypotheses, and an exact source quote. One dynamic decision contract drives the review table, API preview, approval confirmation, and execution payload, so a Google Ads decision cannot silently become a Meta request.
Before execution, the reviewer must select an ad set and explicitly approve. The current build never touches a live advertising account: it creates an idempotent, paused sandbox receipt and stores the audit history locally. Missing budgets, creative, dates, or delivery settings stay visible rather than being invented.
How to test it
The repository includes an Apple Silicon Electron build and a no-credential verification path. From app/, run npm ci, npm test, npm run build, and node scripts/verify-judge.mjs. The verifier checks grounded extraction, paused approval, idempotent replay, and persistent execution history.
Codex and GPT-5.6
I used Codex with GPT-5.6 throughout product scoping, implementation, review, testing, security hardening, Electron packaging, and judge-flow verification. Codex helped turn the initial meeting-summary idea into an approval-first execution workflow, identify hard-coded campaign assumptions, add regression tests, and verify the final macOS package.