I build marketing engines out of AI — and keep the judgement human.

An AI content engine I own end to end — and the one I actually test things on

An English-language site about paddleboarding the Turkish coast, built from nothing: positioning, keyword architecture, the site itself, the writing setup behind it, and affiliate revenue. No team, no budget, nobody to hand a draft to. It is also my sandbox — if a workflow does not survive here, I do not take it into client work.

• One brief file, not one-off prompts. Audience, tone, approved facts, internal linking rules and product criteria live in a single source the whole workflow starts from — so the model never begins from a blank page.
• A prompt library someone else could run. Reusable setups per content type: route guide, gear review, seasonal explainer. Handover-ready, because a system only counts if it survives the person who wrote it.
• QA is part of the workflow, not a mood. Every draft is checked for sources, product specs and claims before publishing. In affiliate content one wrong number costs the reader's trust permanently.
• Built for both audiences. Structured so search engines rank it and AI assistants can find and quote it — and written so a human wants to keep reading either way.
• Interactive where prose is worse. A board-size calculator, a wind-hour timeline and a paddled-routes map, because some answers should be a control the reader moves, not a paragraph they parse.

150+ AI-powered presentation templates, shipped in two weeks

A product launch needed a template library and there was no version of that timeline where a person wrote them one by one. So I built the production line instead: agents generating prompts, descriptions and metadata in bulk, keyword and competitor research deciding what was worth building at all, and me choosing the 24 that would represent the launch.

• Research set the brief. Search demand and competitor gaps decided the categories, so volume went where there was actually a reader waiting.
• Bulk generation, hand-picked selection. The agents made the set possible; the featured 24 were chosen on search volume, design quality and use-case spread.
• Discoverability in two places. Structured with product and design so templates surfaced in search and in ChatGPT-style assistants — before that had a name.

An RSVP apps directory that became a real acquisition channel

I owned this one end to end — what the directory should contain, who it was for, how each entry earned its place, and the copy that made someone act once they arrived. Persona research first, keywords second, because a directory that ranks and converts nobody is just a well-optimised dead end.

• Built on user needs, not keyword volume alone. Research into what people were actually trying to do decided the structure.
• Worked with landing page designers so the page did the same job the copy promised.

One page, position #6, seven paying customers

Permission slip templates: an unglamorous niche nobody was fighting over, which is exactly why it was worth taking. One asset, researched and written properly, still returning customers long after publication — the argument for depth over volume in a single line item.

A Europe and Middle East launch plan — and the claim I refused to make

Inavitas builds smart energy management software across generation, consumption, storage and renewables. The brief was a go-to-market plan for launching its EMS product into European and Middle Eastern markets. The obvious position was hardware agnosticism, because that is what buyers in this category say they want. I did not use it.

The company's own materials describe software, hardware and turn-key solutions together. A hardware-agnostic claim would have been technically fragile, and the product team would have had to walk it back in the first serious engineering conversation. I positioned it as an integration-friendly control layer instead — a software layer above mixed asset environments. Same competitive gap, no promise the product cannot keep.

• Six competitors mapped, one gap named. Fluence, GreenPowerMonitor/DNV, Wärtsilä GEMS, FlexGen and the Chinese hardware-bundled ecosystem — each with its recognised strength and the structural difference that leaves room.
• Three buying dialects, not one message. Finance asks whether returns are protected, engineers ask whether it connects and performs, operations asks whether it deploys cleanly. Each gets its own proof points and its own call to action.
• A launch built for a 6–18 month sales cycle. A quiet foundation of analyst briefings and reference data first, then regionally distinct waves — Europe on regulatory credibility, the Middle East on speed and commissioning.
• KPIs in three layers. Early signal at 0–3 months, pipeline health at 3–9, business outcomes at 9–18 — so the plan can be judged before it is old enough to have results.
• The weakest point written into the plan. If the integration claim cannot be demonstrated against real hardware partners, the message drops back to a weaker one. That is a question for the product team, and the strategy says so rather than hoping nobody asks.

A launch concept for a pre-launch product — including the version I threw away

Upsight automates Upwork bidding, in a category where automation is no longer rare. Competing on speed would have started a fight with a funded incumbent on its own ground. What no competitor claims is proof, and Upsight already publishes it: a transaction-level ledger including refunds and clearing charges, and a funnel with denominators you could defend out loud.

The site is JavaScript-rendered. My first pull returned only the hero section, and I built a defensible but shallow strategy on it. When I rendered the page properly, the product had already said everything I was about to recommend, and said it better. So I threw the first strategy out and rebuilt it on what the product already believed.

• I did not invent a new tagline. The product's own line was sharper than anything an agency would write for it in a week. At launch the senior move is deciding which true claim leads, not adding another one to the pile.
• The copy under-claims relative to the product. Nothing implies human approval protects an account from suspension, because the product's own safety page declines that promise. Copy that out-claims its product destroys the position the product is built on.
• One call to action across everything. A thirty-minute audit of the prospect's real numbers, which they keep whether they buy or not — and two thirds of the effort on founder-led LinkedIn, because the origin story cannot be told by a brand account.
• Assets rendered from code, not assembled by hand. HTML and CSS through headless Chromium at exact pixel dimensions, design tokens sampled from the live product and declared once. Copy changes are one line and a re-render.
• An AI presenter built, tested and cut. A synthetic voice reading a script about honest numbers argues against itself. The comparison was the deliverable, not the file.