Inspiration
A couple months ago one of our friends were looking to buy a car. One roadblock we instantly ran into was the fact that none of us were truly educated about how much we should be spending and where. While using various price comparison tools, we realized that this wasn't just a car purchasing issue, this could apply to many other aspects of our lives where negotiating for a better price could be applied.
What it does
By taking a picture of your receipt or of a document, We strip the data about what was purchased, how much was spent, and where by ZIP code. Then we utilize our own machine learning model to predict the ratio of how much was spent to what a baseline for the area is supposed to be. Then we utilize generative AI through Gemini to create letters tailored to send to the institution you are negotiating with in order to fight for a more fair price.
How we built it
We built SmartSpend by merging open-source price transparency data from the Center for Medicare and Medicaid Services (CMS) with hospital billing records into a unified dataset, which we cleaned and enriched with features like billing codes, payer information, locality, and Medicare benchmarks. To ensure real-world generalizability, we split the data by hospital system using GroupShuffleSplit, preventing data leakage, and encoded categorical variables with LabelEncoder—saving each encoder for consistent inference. We then trained three LightGBM quantile regression models (at the 25th, 50th, and 75th percentiles) to predict the ratio of a given charge to the local Medicare benchmark, enabling users to understand conservative, typical, and aggressive pricing scenarios. These models, along with their encoders, are serialized for integration with our receipt-processing pipeline and Gemini-powered negotiation letter generator.
Challenges we ran into
While we were able to come up with the concept for the model and its use cases, we ran into the problem of user friendliness. While the two of us would be happy to run the program in a simple two button system, we realized most users would prefer something more explanatory and comfortable. But we had a hard time figuring out what was the most user friendly.
Accomplishments that we're proud of
We managed to synthesize and create a finished product! And we couldn't be more proud of that!
What we learned
Sometimes the backend really can be frustrating, but creating a front end with no vision is the most time consuming :(
What's next for SmartSpend
First we want to look in the near future were we apply the same techniques we used to train our model to the utilities and auto financing side of the site. Then we want to build a separate generative tool that allows us to teach other financial literacy skills separate from just better understanding your financial situation. In the long term though, we are looking towards expanding this into a more generalized model that uses private data sets in the future so that everything can be negotiated with a larger block of consumers.
Log in or sign up for Devpost to join the conversation.