The marketing machineA marketing department with nobody in it: AI writes the calendar, renders the reels, posts, measures and learns. Fifteen euros, once.
- 1Six layers of Supabase edge functions run WRDB's marketing with no human in the loop: an LLM writes the content calendar, an AI persona's reels render themselves, publishing posts every fifteen minutes, analytics pull the numbers back and rewrite the strategy. 491 posts on 15 EUR of Higgsfield credits, bought once and not yet used up.
- 2Role: Co-founder: designed and built it. Status: Live.
- 3Stack: Supabase, pg_cron, Meta Graph API, TikTok API, HyperFrames, Higgsfield, Next.js.
- 4Key figure: 15 EUR total spend, once, credits still left.
Nobody in the loop
WRDB has three founders and no marketing budget, so marketing became software. I built a six-layer machine that runs with no human in the loop: a context brain of brand, audience and competitor documents; market intelligence; a strategy layer, an LLM, that writes the content calendar straight into Postgres; a content factory that renders the posts; publishing; and analytics that feed what happened back into the brain. Every layer is a Supabase edge function on pg_cron. Nobody schedules a post, nobody writes a caption, nobody reads the metrics by hand.
The content is made by AI too. The face of the account is a persona who does not exist: her images come from Higgsfield, her reels are rendered from HTML with HyperFrames on a hook bank the machine rotates every day, and each post is tagged by hook style, emotion, call to action, visual style and trend so the machine learns which formula wins per platform. Publishing is a cron every fifteen minutes through the Meta Graph API and the TikTok Content Posting API, one UTM link per post. The only money it has ever needed is 15 EUR of Higgsfield credits, bought once, and there are credits left.
The loop
- 1The brain: brand, audience and competitor documents, plus last week's findings
- 2Strategy: an LLM writes the content calendar into Postgres
- 3Content factory: Higgsfield renders the persona, HyperFrames renders the reel
- 4Publish: every fifteen minutes to Instagram, Facebook and TikTok, one UTM link per post
- 5Pull: daily metrics per post, follower and reach snapshots
- 6Judge: anomaly flags, a composite score per post, A/B winners
- 7Digest: the weekly summary is written back into the brain
| Month, 2026 | Instagram posts | Accounts reached | App signups |
|---|---|---|---|
| March | 31 | 2,417 | 4 |
| April | 36 | 2,551 | 10 |
| May | 106 | 4,600 | 42 |
| June | 95 | 3,601 | 34 |
| July | 28 | 211 | 2 |
| August | 30 | 151 | 3 |
| September, to the 15th | 13 | 36 | 2 |
What the numbers say
Signups track publishing volume. The two months the persona account posted about a hundred times, May and June, brought 42 and 34 signups, more than the previous six months combined. When output fell to about thirty posts a month, signups fell to two or three. That is the strongest pattern in the data and it is only temporal: the attribution pipeline holds 425 tracked links and has never recorded a click, because the Graph API pull does not return link clicks. I report that as a gap, not as a result.
Volume did not buy engagement. The fourteen posts on the founders' own account between December and February reached 11,114 people at 9.5 to 14.5 percent engagement, about 794 people per post. The persona account's 339 Instagram posts reached 13,567 at under one percent from April on, about 40 per post. The machine wins on cost and loses on resonance, which is the case for pointing it at proven hooks rather than at more volume.
Cost is the easy part. Fifteen euros of credits, bought once and not yet used up, is everything the machine has ever cost: 0.03 EUR per post, 0.61 EUR per thousand accounts reached across both accounts. Since the persona account went live on 7 March the app has gained 97 signups; counting them all as the machine's gives a CAC of 0.15 EUR per signup. That is a ceiling, because the attribution pipeline never recorded a click and the link is temporal, and it is also the only kind of CAC a zero-budget channel can have. Cash spent on distribution is zero. The real cost is the human hours it does not consume.
Review of
Formulas
- C
- is everything the machine has cost: 15 EUR of Higgsfield credits, part of it still unspent
- R
- is accounts reached on the persona account, 13,567
At most 1.1 EUR per thousand accounts reached on the persona account, and 0.6 EUR across both accounts. The credits are not used up, so the true figure is lower. The plan's target was under 1 EUR.
| Metric | Target in the plan | Actual |
|---|---|---|
| Content cost per thousand reached | under 1.00 EUR | at most 1.1 EUR persona, 0.6 EUR both accounts |
| Cost per view | under 0.001 EUR | at most 0.001 EUR |
| Cost per signup | under 0.50 EUR per install | at most 0.15 EUR, 97 signups since 7 March; 0 clicks attributed |
| Engagement rate | over 5% | 9.5 to 14.5% founders' posts; under 1% persona |
| Paid distribution at this phase | 0 EUR | 0 EUR |
| Human hours per post | 0 | 0 |