AI Creative Origination & Production Test
+
You already know us
This isn't the first thing we've built for Samsung.
{{ c.title }}
{{ c.body }}
So when the test landed, we didn't start from zero. We mobilized the Flywheel team and ran it through the model we are building for exactly this.
AI in production
But first: Remember this
AI origination test.
We made this film for Samsung, start to finish, with AI. A like-for-like render of the quality this team ships. Scroll on whenever you like.
How we approached the test
One source. One system. Two chapters of work.
You gave everyone the same assets and the same deadline. Here's what we made, and, just as you asked, exactly how.
Chapter 01
01The CRM.
One Back-to-School concept, assembled into seven personalized emails from a single modular template.
The work · 7 emails
One template. Seven audiences.
Tap a segment. Watch the email reassemble.
{{ seg.subject }}
{{ seg.preheader }}
{{ seg.personLabel }}
{{ p.k }}: {{ p.v }}
Assembly
Deliverability
View deliverability report↗Mailgun seedlist test — inbox placement, spam score and client rendering.
Deliverability
Deliverability report — pending↗Dynamic where it counts. Fixed where it should be.
Chapter 02
02The Video.
One 60-second master. Localized, cut down, character-swapped, and delivered in every aspect ratio.
The source
It all starts with one master.
The deliverables
One master.
24 renditions.
Hours & process
20 is not 80.
The point of this test wasn't only the work. It was the how. Here's the full accounting: hours by role, tools per task, and how the work was made.
Appendix · Independently attested
The receipts.
Production & QA version history · Flywheel × Samsung · AI Creative Production Test · July 7–9, 2026
Why the receipts start here
Creative development (strategy, generation, layout) is documented in this submission the way the brief asks: hours by role, tools per task, and revision rounds per deliverable. Those numbers are ours, and we stand behind them.
But from the moment an asset was locked for production, we stopped being the only witness. Every step below is timestamped by a system we do not control: GitHub records every commit, Mailgun records every rendering test, Vercel records every deployment, and the CDN records every cache purge. None of these records can be edited by us after the fact. This appendix is that machine-written history: the last mile of production, independently attested.
The same principle governs our deliverability reports: every email variant links to a live, third-party report on Mailgun's own domain. We chose verification we cannot curate.
Upstream creative history (Figma version history, Canva edit logs, Adobe cloud document versions) exists inside those platforms and can be walked through on request in a working session.
Timezone note: GitHub timestamps are UTC. Mailgun timestamps are as displayed in the dashboard. Sequences are consistent across both systems.
Sources of truth · live, third-party
Ledger A · Email asset history (samsung-email-test, UTC)
Full Samsung Sharp Sans family (regular / medium / bold, official TTFs) committed within #3–4 window alongside image fixes.
Ledger B · Rendering QA history (Mailgun Inspect)
Every test below is a full multi-client rendering run (up to 104 clients on the full matrix; 65 on the Samsung 26 Test Suite profile). ~1,500 individual client previews were rendered during this engagement (account meter: 2,005 total).
Original creation times were re-stamped when these tests were archived to the Archive folder on Jul 9 · 04:42 AM. Session records preserve the sequence:
The residual "QA 2" on every final test is the same two samsung.com pages (/computers-and-monitors/, /registration/) that block automated link-checkers while loading normally in browsers; verified manually, constant across all variants.
The trend the ledger shows
First variant: 5 iterations to clean. Middle variants: 2–3. Final variants: clean on the first run. The system learned: every defect class discovered was folded into pre-flight checks for the next variant. That is the compounding the deck describes, visible in third-party timestamps.
Ledger C · Presentation build & deploy (AI-creates, UTC)
Every commit above triggered an independently logged Vercel deployment.
What QA caught before the client ever saw it
Timestamped within the ledgers above
Dead samsung.com URLs across variants, including Samsung Canada's own student-discount page returning 404 during a back-to-school campaign. Found, correct destinations researched and verified live, links re-pointed.
Image export artifact: a single-pixel white keyline on the hero KV, invisible on light backgrounds, visible in dark mode. Fixed at the pixel level and re-shipped through the CDN within 40 minutes of the first test (21:38Z → 22:18Z).
WCAG AA contrast failures across two variants' palettes, remediated hue-faithfully, every value documented (e.g., 2.3:1 → 5.79:1).
Outlook rendering defects in every VML overlay pattern (text clipped or floated in Outlook Windows), root-caused (unpinned frame heights) and systematized into the template.
Accessibility to zero: lang attributes, heading order, link affordance, alt-text uniqueness. Every variant shipped at 0 accessibility issues on a live third-party report.
Where this goes
Two days for the test. Imagine the year.
Same brand bar, held by people. Same system, getting faster every turn. That's what production at scale looks like with Flywheel.
+