At 3am, nobody reads a forum thread

The spark

Mylo is a parenting app with active community forums. The forums worked. That was the problem. Parents were getting good answers out of them and paying for it in time they did not have, scrolling long threads at the exact moment a baby was unwell or would not sleep.

Underneath that sat two more pressure points. Sleep data, growth milestones, medical records, and photos were spread across separate apps, notebooks, and camera rolls. And generic parenting advice rarely matched the actual stage, temperament, or circumstance of the parent asking.

Bolting an AI assistant onto a parenting app for novelty was the obvious move and the wrong one. The brief we accepted was narrower: anticipate what a parent needs when scrolling is not an option.

The dig

In-depth interviews with current Mylo users, plus an analysis of how those same people used competing parenting apps and forums. Specifically what they reached for first, what they abandoned, and where the tools ran out.

Four things came up in every single interview. Real-time support mattered most in the middle of the night. Community trust outweighed expert authority, because parents believed other parents who had been through it over a well-sourced article. Tracking felt like unpaid work unless the data showed something back. And memories deserved a deliberate home rather than the chaos of a camera roll.

The second of those is the one that changed the product.

The shift

We had scoped an AI feature. Research reframed it as a corpus feature.

If parents trust other parents more than they trust expert content, then the valuable thing Mylo owns is not a model. Anyone can call a model. It is years of community answers from parents who have already been through the exact stage you are in. So Mylo GPT draws on the app’s own forums and chats, and its answers can carry that provenance. That is a defensible product, and a generic assistant trained on the open internet is not.

The same reframe set the bar for the tracking features. It became a rule: no tracker ships unless logging pays the parent back. Sleep tracking earned its place because patterns emerge. Growth data earned it because projections do. Anything that only produced a tidy record got cut.

The valuable thing was not the model. It was the archive of parents who had already been there.

The build

Mylo GPT, a personalised assistant answering from the community archive, plus four supporting features that each cleared the payback test: a sleep tracker, a baby profile with growth stats and projections, a medical record repository, and a memorabilia space for milestones.

Provenance list · below every answer

The design system carried Mylo’s existing warm corals and sunlit yellows. Those already meant something to existing users, and throwing them out would have cost recognition for nothing. What we added was structure: a typographic hierarchy, an icon set, defined button states, and the interactive pieces that were new: the feedback slider, the sleep tracker bars, and the activity ring.

The point of the system was that a future feature could be built from it without breaking the feel of the app.

Seven-night window

Constraints

Fourteen weeks, a team of three, and academic work, which sets a hard ceiling on what can be claimed. There is no install base and no retention curve.

Mylo GPT was designed against a corpus we could read but not query. We had no engineering route into the live forum archive, so retrieval quality, answer latency, and what happens when the archive has nothing useful for a given stage are all unanswered. A provenance list is only trustworthy if the retrieval behind it is, and we could not test that.

Medical records were scoped as a repository and left there. Handling real paediatric records properly is a privacy and compliance problem, not a UI problem, and fourteen weeks was not the place to pretend otherwise.

The proof

Fourteen weeks, ending in usability testing on high-fidelity prototypes. The signal is qualitative.

The assistant tested as useful specifically because answers were traceable to other parents, which is the reframe holding up under someone else’s hands. And the four supporting features survived the payback test one at a time rather than being justified as a bundle, which is the harder thing to get right.

The lesson

I would test for the payback earlier. We arrived at “a tracker has to give something back” in synthesis, after four features were already sketched, and one of them nearly shipped on the strength of looking complete. That question belongs in the interview guide, not in the analysis.