Thought Leaders in Artificial Intelligence: Zohar Bronfman, CEO of Pecan.ai (Part 2)

Sramana Mitra: I’m going to back up a moment and ask you to describe your process of arriving at that collection of use cases. What were the drivers? What was the process of arriving at that?

Zohar Bronfman: On one end, we wanted something that is rich in data, easy to connect to the business need, and relatively easy to translate into non-data science terms. On the other end, we looked for where we thought the market would see the fastest time to value. Many times, you have amazing data science use cases but the whole cycle of value realization can take even two years. We wanted a platform that allows us to quickly and unequivocally show our customers that they have tangible value.

One of the illnesses in the world of data science and AI platforms is that the technology is amazing, but the tradition is overwhelming. Eventually if you don’t tie it to value, you are in the risk of ending a nice-to-have instead of something that the business must have. 

Sramana Mitra: What’s in the platform?

Zohar Bronfman: There are mainly two components which help us realize the vision I mentioned. Let me start with the second component. It’s the AutoML component. You get an AI-ready dataset and based on this, you automatically build many predictive models.

The first component is our data blending and data structuring component, which is where we would connect as a platform to the raw data sources that a customer has. It can be the customer table, purchases table, or products table. We will connect to those tables on one end. This first component will take all of those tables along with the input that we’ll get in the UI and automatically create the AI-ready dataset which is the input to the second component. Our core competency is our ability to automatically generate the AI-ready structure from a set of regular transactional and behavioral tables that the customer or the user possesses. 

Sramana Mitra: What is the difference between regular transactional data that you are pulling from existing databases of customers and the AI-ready data structure?

Zohar Bronfman: Let’s take a simple example. Let’s say you have, on the one hand, a customer table which is a table that describes data about the customers. For example, the customer ID, zip code, age, gender, email. Then let’s imagine that you have a table that is the basic transactional table where you have the transaction ID, type of transaction, maybe a purchase amount and date.

These two tables are tables that you can’t do anything with regards to ML or predictive analytics. We said you had to transform them into something that is meaningful. You have to create an entity that has features and labels that stand for an event that you would like to learn from. We would join or blend these two tables and generate a sample that represents the historical transactional behavior of a specific customer over the last two years for example.

That would mean that we would go to the transaction table, identify all of the transactions that were generated by the same customer ID, and take the timestamp of those transactions. Based on the timestamp and based on the customer ID and data, we are able to generate a sample for the machine that will have a few data points about the customer and also the sum of the last five transactions or maybe the time since the last transaction.

The platforms compute automatically and aggregate and generate for this sample that represents the customer behavior. All of that is done automatically. Eventually, you’ll have the features per entity; entity being the customer at a specific point in time and features being the variables that we’ve aggregated based on various time-based static tables. 

This segment is part 2 in the series : Thought Leaders in Artificial Intelligence: Zohar Bronfman, CEO of Pecan.ai
1 2 3