Inspiration
We're in 2023, computing has arrived to a level that even humans don't know what is real and what is not. Based on that, we thought to give a turn to Restb.ai API, that uses AI to classify real estate images in order to extract features from them. Using their API we tried to send not real images, but 3D generated plan (a frame or some frames) in order to extract those features before even the first brick is placed. Knowing that, the user can make changes to their ideas (appliances, number of rooms...) and know how much those changes affect the market price for their home.
What it does
Given different frames from a 3D plan, we process them and extract the different features that it contains and we also calculate how many rooms are in the house. Using this information and a Machine Learning algorithm joined with the knowlege of where it will be built the house we're able to predict which will be the market price of them.
How we built it
We started working in parallel on both the webApp and the processing and making the machine learning algorithm and then, once we had obtained a model (not necessarily the final one, but one that has the same imput and the same output as the final one), we could test it on the web. As soon as these tests gave results, one part of the team looked for a way to improve the algorithm and the other one started with the telegram bot to offer our service through it. Finally, at 5 o'clock in the morning we got everything working and decided to take some rest.
Challenges we ran into
For the Machine Learning algorithm, we started assuming the data was clean (we assumed it was data from Restb.ai and at least checked). But once we started cleaning and checking the data we saw that it wasn't, we discovered that even the values were missclassified between city and neighborhood.
WebApp, We had no idea how to use a machine learning algorithm, so we had to learn what it was and how to feed the data into the trained model. Parsing the data received from the different APIs it's a been an uncomplicated but entertaining task.
Accomplishments that we're proud of
- Developing a Machine Learning algorithm with +80% of accuracy in our first real data.
- Developing a WebApp using NodeJS without any previous experience including AJAX techniques.
- Our first working Telegram bot.
What we learned
On one hand we learned how difficult it is to read raw data and try to clean it in order to have a good dataset to train a model. Talking about models, we also tried different models and checked which was the best and why. On the other hand we also enjoyed the time together, big moments of fear when we didn't know how to continue or what solution we should keep but at the end, all of those moments ended with a lot of enthusiasm when we succedded.
Talking about skills related to programming, we really learnt a lot of things, as we said before, it was our first time using real data (using 300k rows) to train a model and have a "good" accuracy. Web is also a new thing for us, watching tutorials on the go to solve the different doubts we had.
What's next for ReAI Estate
Maybe Restb.ai hadn't thought about this 3D part and we could join their team and continue thinking outside the box with this amazing team that helped us understanding the challenge during those days (and nights ;p ).
Log in or sign up for Devpost to join the conversation.