AI education · content and outbound systems · AI Lab, iPhone Photography School · March to September 2026

An AI education channel run as a system, then an outbound machine built in 3 weeks and instrumented so the data could say no.

Two systems for one brand: an idea-to-published-video pipeline for the tutorial channel, then a cold email machine that told us, with numbers, that the channel was wrong for the buyer.

4,349emails to 2,661 businesses, a 2.2% human reply rate, from a fleet grown 2 to 10 domains and 104 to 320 mailboxes in 3 weeks Campaign export, 4 August to 7 September 2026.

~10hrs/wk saved on the content side across a split team, with a person approving every send

At a glance

Client: AI Lab, the AI education brand inside iPhone Photography School (a 1M+ audience company, Latvia). Contract, March to September 2026.
Who uses it: 2 educators, an operations lead, and a leadership team reading the weekly numbers; members watching the tutorials.
Shipped: March to September 2026. 6 months, in two phases.
Stack: Claude Code
Codex
MCP
Python
Wistia API
Circle
Customer.io
Smartlead
Clay
Apify
Google Workspace
Slack
Kie image API
Whisper
Cost to run: About $1 per 30 researched leads. The content pipeline ran on existing subscriptions.
My role: I built all of it as the in-house AI engineer, alongside a second educator on the content side. The sending stack was inherited from a departing contractor and rebuilt.

What I built

The build

Content: an idea engine with semantic dedup against every video ever made, mandatory research, and a scoring rubric with hard gates; a taste model mined from months of accept and reject decisions that cut the context loaded per run by 79%.
Content: a one-command publishing pipeline with a loudness gate at a fixed -14 LUFS, Wistia upload into the right series, tracker write, thumbnail set; a Monday, Wednesday, Friday newsletter builder that stages a verified test send for a human to approve; a weekly KPI report leadership reads.
Outbound: sourcing from Google Maps and public registries, Claude research writing one factual opener per business, an automated rule gate on every email, and 2-step plain-text sequences from 320 mailboxes on 10 domains under capacity rules, at about $1 per 30 leads.
Offer and ICP: 11 niches scored on 2026 data and cut to one buyer with a kill date; 2 complete three-tier offer chains; a savings calculator; the delivery playbook for a workshop product; a clean two-day wind-down with a full export and a restart guide.

What changed

Before and after

Before

By hand

Deciding what to filmA gut call and a spreadsheet

Publishing a videoManual upload, manual tracker

NewslettersWritten by hand 3 times a week

Sending capacity2 domains, 104 mailboxes

What the data saidUnknown

After

The system

Deciding what to filmScored ideas that pass hard gates

Publishing a videoOne command, loudness-gated

NewslettersBuilt, staged, approved by a person

Sending capacity10 domains, 320 mailboxes, warmed

What the data said2.2% human replies; the question beat the pitch every A/B; not the channel for this buyer

Where the numbers come from. Outbound numbers are from the campaign export for 4 August to 7 September 2026: 4,349 emails, 2,661 businesses, 91 replies of which 51 were human (2.2%), 3 marked interested, 0 confirmed calls. That last number is the point: the system was built to report it honestly, and the restart guide says why. The ~10 hours a week on the content side is the team's own estimate.

For engineers

How it's built

Architecture: A shared context layer (voice, rules, trackers, decisions) synced to both Claude Code and Codex, so any agent picks up where the last one stopped. Skills for each job: video ideas, post-video, newsletter build and publish, KPI report, lead enrich, lead check, email capacity. MCP and API connections into Wistia, Circle, the email platform, Google Workspace, and Slack.
Where it runs: On the educators' machines, with a person pressing start. Sending ran on Smartlead over Microsoft mailboxes; sourcing on Apify and free registries; the KPI report published to a stable link.
Tests and evals: A rule gate on every cold email (no years, no street names, no clinical language, no 'AI', sign-off matches the inbox), run as a script before any batch. A lint on the KPI report before publish, with a corrections trail when a number was wrong. Every capacity number came from the script, never from memory.
Cost and latency: About $1 per 30 researched leads. The content pipeline ran on existing subscriptions.
Guardrails: A human approves every send: newsletters stage a verified test, campaigns stay drafts until a person activates them. Capacity rules (2 emails per inbox, 100 per domain per day, 14-day warm-up) enforced in code. The whole outbound system was wound down in 2 days with a full export and a restart guide.
The tradeoff we chose: Why this way and not the other
The opinion question beat the direct pitch in every A/B pair, so we stopped pitching. And when the numbers said cold email was not the channel for this buyer, we reported that instead of optimizing the copy again. The system's job was to make that finding checkable.
Code: The live proof pages (offer chains, ICP decision, KPI report, savings calculator) are still up on Vercel; links on request.