# Two AI features I built: one for pet parents, one for home buyers

> Lorem answers questions in the Pet Parent App and Roost analyzes deals on Roosthaven. The two products share nothing, but I built the AI part almost the same way both times.

URL: https://www.mnthan.com/writing/shipping-ai-two-products
Published: 2026-05-08
Author: Manthan Thummar
Tags: AI, Product, Architecture

I have built AI into two products. They have nothing in common.

Roost is the co-pilot on Roosthaven, a real-estate platform I built alone between December 2024 and April 2025. It reads disclosures, contracts, inspection reports and offers for people who are buying or selling a home.

Lorem is the companion inside the Pet Parent App at Happy Pet Tech, where I am CTO. It answers questions from pet parents about health, diet, training and behaviour.

One is about the biggest purchase of someone's life. The other is about whether a dog can eat grapes. Still, I built the AI part almost the same way both times, and it is easier to show that side by side.

| | Roost | Lorem |
|---|---|---|
| Who asks | Home buyers, sellers, agents | Pet parents |
| What it reads first | The deal: listing, tours, offers, uploaded documents | The pet: species, breed, age, weight, history |
| Where it runs | Behind the platform's API | Behind the app's API |
| The answer it must not give | A general opinion about houses | A guess about a sick pet |

## What it reads first

Look at the second row. In both products, most of my time went into that row and not into the AI model.

Roost is useful because it has read the documents of this deal. When it compares two offers, it compares the real offers. Lorem is useful because it has the profile of this pet. The work in both cases was deciding what the model gets to see before it says anything. That is plain backend work: fetching the right records and passing them in.

Without that row, both features are worthless. Nobody should decide on a house based on what a model thinks about houses in general. And pet parents can find general advice anywhere.

## Where it runs

Both features sit on the server, behind the product's own API. The clients are thin.

On Roosthaven this meant Roost could use everything the platform already knew, without the front end carrying any of that logic. In the Pet Parent App it matters even more, because the client is a mobile app. We change Lorem's prompts and safety rules often, and none of those changes wait for an app store review.

## The answer it must not give

Both products can do damage with a wrong answer that sounds sure of itself.

With Lorem, some health questions have only one responsible answer, which is to go to a vet. Lorem is built to say that. With Roost, there is real money involved, so its analysis stays tied to what the documents say and does not wander into opinion.

In both cases I decided these limits at the start, before the feature was built. I did not add them later as a patch.

Roosthaven has since been exhibited at C.A.R. REimagine 2025, the California Association of Realtors' conference. The Pet Parent App has passed 12,000 downloads.