Apartmatch
Easy, personalized apartment hunting.
Inspiration
As students, we've all faced the dreaded dilemma of finding a place to live next year. Hours wasted scrolling through Facebook, Kijiji, Zumper... and then summoning the courage to chat with real estate agents - it's a nightmare that should be reserved for horror movies, not house hunting! Reading each description and browsing through 20+ photos to weed out potentially problematic apartments is time-consuming and annoying 😡. Imagine if you could spend that time cramming for finals instead!
What it does
ApartMatch gathers loads of listings into a simple, sleek interface. You just input your budget, the number of bedrooms and bathrooms you want, and a short text description of your dream apartment, and we work our magic. Using image captioning, we analyze the listing photos to provide you with a more truthful description of the property, because let's face it, that "beautiful third bedroom" doesn't always come with windows! We then combine these enhanced descriptions with the information from the original listing and employ advanced semantic matching techniques to sift through the options and provide a curated selection of the best apartment matches for you.
How we built it
Our Flask-powered backend allowed us to seamlessly integrate our machine learning models into the website. We fine-tuned a pre-trained HuggingFace vit-gpt2-image-captioning model to analyze the listing photos. We then also used a pre-trained PyTorch bi-encoder to vectorize the user's input description of their ideal apartment as well as the descriptions listed on the website. The model then computes cosine similarity scores to determine how closely a listing satisfies the user's dream apartment requirements and displays the top 10 matches. We built the front end of our website using Bootstrap, ChatGPT, and lots of tears.
Challenges we ran into
We frequently ran into rate-limiting and authorization issues with the realtor.ca API, and had to spend a lot of time debugging it.
Accomplishments that we're proud of
We are thrilled to have built something that we can actually use next year when finals season meets apartment-hunting season.
What we learned
This was the first time we had the opportunity to work with NLP techniques. Throughout the hackathon, we tested a few different models to generate captions and do semantics matching. We were able to gain a better understanding of how each of them worked and how to choose one that was optimal for our use case. We also tried to integrate a pre-trained housing price estimation Random Forest model to indicate whether or not a given listing is a fair price but ran out of time before it was finished. Finally, as all of us were relatively inexperienced at web development, we learned a lot and gained a tremendous amount of respect for front-end designers.
What's next for ApartMatch
We would like to refine the sorting algorithm and include more websites. For our prototype, we limited ourselves to just realtor.ca, as it would have been very time-consuming to consolidate the listings from multiple different websites, but it would be an excellent addition to further streamline the apartment hunting process. Additionally, we would like to work toward creating a more cohesive UI design and UX flow, and add even more features. For instance, we want to show all of our listings on a map to allow users to more easily assess an apartment's proximity to amenities, transportation, and other key factors. 🚀 🏡