Within about a year I built AI into two products that have nothing to do with each other. Lorem is the companion in the Pet Parent App; it answers pet parents’ questions about health, diet, training, and behaviour. Roost is the co-pilot on Roosthaven; it analyzes disclosures, contracts, inspection reports, and offers for people buying and selling homes. Different domains, different users, different stakes. The engineering lessons came out nearly identical.
In both products, the model was the smallest part of the work. Lorem is useful because it reads the pet’s profile and records before answering, so “is this safe for my dog?” gets an answer about that dog. Roost is useful because it reads the deal’s own material: the listing, the tour activity, the offers, the documents people upload. Both times, the real design work was deciding what the model sees before it says anything, and that work looks like data plumbing rather than prompt cleverness.
A generic answer is worthless in both domains, for different reasons. Pet parents can get generic advice anywhere; they open the app because it knows their pet. And nobody should weigh a house purchase on a model’s general opinion about houses; the analysis has to come from what this deal’s documents say.
Both AI features live server-side, behind the product’s own API, with thin clients in front. In the Pet Parent App that decision earns its keep constantly: prompts, context assembly, and safety rules keep changing as we learn what people ask, and none of those changes wait on an app store review. On Roosthaven the same shape meant Roost could draw on everything the platform already knew, listings and tours and uploaded documents, without the client carrying any of that logic.
The two products taught the same caution from opposite directions. With Lorem, some questions cross a line where the only responsible answer is “take your pet to a vet”, and the feature is built to say exactly that. With Roost, the money involved means the analysis stays anchored to the documents instead of drifting into opinion. In both products the failure that matters is a confident wrong answer, so the limits were designed in from the start.