AI Product Manager who builds
Ten years in product. I take equal ownership of what ships and of whether anyone uses it, and I build the thing myself when that is the fastest way to find out.
The problem. Writing down how a company actually works takes weeks of consultant time, so most teams never do it, and everything downstream of that document stalls.
What I did. Built an agent that interviews the person who runs a process and returns a structured work instruction. Taken from first idea to production on Claude Code over a Supabase and Postgres backend. I also set up and run its paid acquisition, so the funnel is mine as well as the product.
The obvious build was a form with AI-generated text. I rejected it: people cannot describe their own process in a blank field, which is exactly why the document never gets written. So the product interviews instead of asking, and the quality bar moved from "did it produce text" to "did a real person get a usable instruction out of it". I chose activation as the measure, a finished instruction, rather than revenue, because revenue this early would tell me about my ad spend and not about the product.
The problem. A pet telehealth marketplace with consultations selling one at a time. Nothing held anyone beyond a single visit, and the supply side was thin.
What I did. Took the Care Plan subscription from concept to market in six months, then stayed with it: opened new specialist verticals to grow supply, extended consultations from app to web as a new acquisition channel, built a referral programme and rebuilt onboarding.
Trial-to-paid was the number that decided whether the subscription survived, so it got the experiments rather than the homepage. The one I remember is a 50% first-month discount: it lifted conversion, and the real question was whether those people stayed past the discounted month or just arrived cheaply. That is the kind of result that is easy to report as a win and harder to defend, which is why I would rather talk about it than list it.
The problem. Patients could not find care on the site. The plan on the table was a better catalogue.
What I did. Watched how people actually searched and moved the work to search and discovery instead: rebuilt search on Elasticsearch, tolerant of typos and the wrong keyboard layout, restructured the architecture across services, doctors and clinics, linked 1,000+ doctor profiles, added geo-personalised content across 20+ cities, and shipped booking and schedules integrated with the clinics' medical information system.
A catalogue assumes people know what they are looking for. They did not. They typed a symptom, a doctor's surname with a spelling error, or a district. Shipping the catalogue would have produced a tidier version of the same failure, so the harder conversation was the one worth having. The numbers above are the result of changing the question, not of building faster.