Most AI products get built the same way…

Note · 25 Aug 2026

Most AI products get built the same way: take a language model, wrap a chat UI around it, ship it. For a parenting app that would've been the fast route too. We built Julie backwards instead.

I've been advising on, and building, the tech behind Julie AI, a parenting companion for the first three years of a child's life. Early on we had to answer an important question: what happens when a parent asks Julie something at 3am about their newborn's fever, and the model just... makes something up?

So the language model isn't the star of the system. It's the fallback. The core is a RAG system querying a curated knowledge base built from NHS and NICE guidance, reviewed by midwives, health visitors and sleep consultants. The model's job is to hold the conversation, not to be the source of truth. If the knowledge base doesn't have a grounded answer, Julie says so, rather than confidently inventing one.

It's slower to build this way. Every entry in that knowledge base is hand-curated, and it shows in the pace of the roadmap compared to "just prompt it better". Parents are already asking ChatGPT about their baby's sleep at 3am whether we like it or not, and the answer they get back is built for the average baby, sourced from wherever, with no way to check its working.

Trust doesn't come from a good UI or a friendly tone. It comes from being able to trace an answer back to something a clinician actually signed off on. That's a much harder product to build, and I think it's the right one.

Curious where else this pattern applies — knowledge-first, model as a last resort rather than the main engine? Feels relevant well beyond parenting.

Reply onLinkedIn

More notes

See all →

Last week Uber and Pony AI announced plans to put over 2,000 robotaxis into four European cities.

At the beginning of this year, the UK lost around 2,500 shared cars, not to a technology problem but to parking permits, insurance costs and congestion charging.

I don't think those two facts are unrelated. I think they're the same underlying issue.

20 Aug 2026 Continue reading →

Zipcar's UK exit took nearly half of the national car club fleet off the road. In London it was even higher, over 90%. That's around 2,500 vehicles gone in a few months.

It's easy to read that as a failing business model. It isn't — the model works. CoMoUK's research from May puts the saving at roughly £2,800 a year versus private ownership, with each shared vehicle replacing about 23 private cars (28 in London).

18 Aug 2026 Continue reading →

After more than a decade at Hiyacar, and with a heavy heart, I've made the difficult decision to move on.

I'm immensely proud of what we built together over the years. I’m especially grateful to the incredibly passionate & curious grafters, both present and past, who showed up every day and did exceptional work often under challenging circumstances. Every single one of you made it worthwhile.

21 Jul 2026 Continue reading →