KnotIQ

JetBlue customer insights
JetBlue generated ad

Inspiration

When we heard about the vast transaction data Knot collects, we were reminded of a class we took with Wharton Professor Pete Fader, who's life's work revolves around capturing CLV from aggregate data. We wondered what would happen if we applied his teachings to this individuals' data available with Knot.

What it does

KnotIQ gives merchants their most potential high-value customers and a foolproof, easy starting point to generating and deploying ads based on customer profiles so they can invest in advertising where it will be the most successful, period.

How we built it

We built it with a Python backend, using Numpy and statistical models to generate CLV estimates and customer segments, and then OpenAI to consolidate the segments into a readable format. We used Modal's stable diffusion model for the ad generation, and a python script to generate mock customer data based on the samples given to us by Knot.

Challenges we ran into

Generating data that was somewhat reflective of real customers was the first challenge. We wanted to create data that wasn't already segmented in a pre-decided way but also not completely random so that we couldn't draw any conclusions. What ended up being the most specific was a method that predefined probabilities of certain transactions taking place at different intervals. The second challenge was our actual methodology in segmenting customers. We started with K-means clustering and Regression trees to cluster customers the most similar to one another when we realized that merchants aren't looking to just group customers relative to each other, they actually have to target them. So instead, we decided to apply some marketing formulas to the dataset and find the best customers for a merchant as opposed to the customers most like one another.

Accomplishments that we're proud of

We are proud of having learned to integrate and learn about new technologies (Modal, Knot). We are also proud of how we dealt with the lack of data and were able to strategically generate our own to serve our purpose.

What we learned

We learned that it's really important to visualize and clearly define goals before jumping into building straightaway.

What's next for KnotIQ

We hope to be able to apply this to real data, generating real insights and feedback to be able to refine the customer selection algorithm. For example, we want to be able to calculate a probability that a customer buys from Merchant A, and see that prediction through with real time data.