Real Estate Gems

Landing page
Navigation Map, with pointer moving on click and interactive landmarks.
Dropdown menu with types of houses to search by
Pop up after submitting an example query.

Inspiration

Going into this, we wanted to build something while learning a new set of tools. It's the first hackathon for two of our members, and the third was heavily sleep deprived and only awake because of copious amounts of caffeine. As one does, we examined our weaknesses and decided to build a project consisting of exactly that.

What it does

Real Estate Gems (tm/p/c/etc) utilizes the vibes based decision making by Google's Gemini Pro 1.5 (broke college student edition), paired with the Melissa API and the soon to be deprecated Realty Mole API. In order to aggregate properties that are listed on the market, we pair the basic information about the addresses that have current listings from Realty Mole API with the more complete data set provided by Melissa API, feeding that into a model which has been trained to select the location that it believes to best suit the user's desires and preferences.

How we built it

We split the project into three parts: the AI model, the frontend with the maps and input, and the backend that handles incoming and outgoing API requests. For the backend we settled on a RESTful API structure using express.js, which would take every submission and aggregate the housing data before bundling them up to be sent to the AI model. Our frontend comprised of React library components (and some Chakra UI components), sitting on top of the basic React app that uses the axios library to communicate to the backend, and the React Google Maps API to render the interactive map component. The model is a Gemini Pro 1.5 trained for structured output by a whopping 20 handwritten test cases.

Challenges we ran into

One of the first challenges we encountered was before the project even began; our original fourth member got sick on Friday morning and had to drop. They were also the most technically advanced developer when it comes to frontend design and development, but we pushed forwards anyways.

Our next challenge was our AI model; training it with data was a hassle because the data needed to be assembled by hand -- we had a very specific set of inputs made by combining the data of the two APIs, and when Ethan asked the resident grunt worker ZOTGPT to do it for him it failed miserably. So he made the brilliant decision to handwrite everything and not ask his teammates for help.

Accomplishments that we're proud of

Learning a new stack and experimenting with different APIs and having a good time!

What we learned

APIs are insanely expensive and it's very easy to lose a good chunk of tokens while attempting to debug an interaction with a third party API.

Read Google Gemini's documentation before choosing it based on price because they are astoundingly far behind their competition when it comes to the capabilities of the models.