Sramana Mitra: What kind of customers were you working with before you raised the money? You mentioned that you built your business in a bootstrap mode, and then demand for your company or products went up. That is when you decided to move to Silicon Valley. Tell me a bit about what happened in the pre-funding incarnation of the company. Who were the early adapters of this kind of technology?
Omer Artun: The commonality between all of our customers is that they are all in a high-volume marketing and sales environment. This means there are millions of transactions, thousands of products, and billions of clicks. Companies like Sports Authority, Shazam, and Ideeli, are the type of companies we are working with.
SM: So B2B retail is mostly your segment?
OA: I wouldn’t call it retail. Any time you have a direct transaction with a customer, whether it is B2C or B2B, we can help you. Would you call Shazam a retailer? No. We have a large PC manufacturer, [the name of] which we can’t disclose, as a customer. We are helping with their B2B efforts. It is not a retailer; it is a manufacturer that sells directly. Any time you have a lot of customers and a lot of interactions with those customers, that is where we come in. Let’s take a seller of X-ray equipment, for example. Let’s say they sell 500 [pieces of equipment] to 5,000 hospitals. That is not our core focus.
SM: So, your customers need to have lots of customers, whether it is B2B or B2C?
OA: That is right. We believe that when you have lots of customers and those customers mean having lots of interactions, mathematics and machine learning are really needed. This way you can cut through the noise and understand information within this large data set. This doesn’t mean you will get to know detailed, one-on-one information, though. As an example, if you have 15 bills at any given time, you can know all those 15 bills in detail. But if you have millions of visitors to a website, you can’t know all those customers or what they are looking for. In that case you need a more automated system which will help you distill this information.
SM: Let’s talk about specifics. We have seen applications for machine learning like collaborative filtering algorithms in Amazon and other companies that do online recommendations. I assume you are doing something more sophisticated than that. My own background involves IT, too, so I would love to hear a little bit more about what you are doing on an algorithmic level, without disclosing your “sweet sauce,” of course.
OA: Collaborative filtering is definitely one of the tricks in our bag, and we utilize it. But we also have other algorithms, for example, clustering, which we use to create mixtures of expert-type models; whether it is collaborative filtering or propensity modeling. We use clustering a lot, and we use it in different contexts to solve different problems. If you are selling shoes as a retailer, for example, you might say: “Every person buys 5, 10 or 20 different pairs of shoes. But if I look across, can I categorize these people, in a self-learning way, into different groups that behave similarly?” Then you might encounter categories such as businesswomen buying business shoes and also sneakers for their kids. Then you can have athletes buying athletic shoes most of the time, but also buying other shoes. You can start grouping people together based on their product behavior. Our system does that automatically. It doesn’t mean that the businesswoman never buys sneakers; it just means that her general interests are business shoes and shoes for kids. We create those groups. When the marketer then sends an email, instead of sending out one email with merchandise for everybody, they can start contextualizing the email or the website for a specific person. Now you are not sending 100% one-on-one emails anymore, but you get to a point where you are relevant to the audience.
SM: Personalized recommendations for customers.
OA: That is right. Think of it like in the collaborative filtering example. If I go to Amazon looking for audio amplifiers, everybody would get the same type of audio amplifier when accessing the same audio amplifier category page. Underneath that I see product recommendations specifically made for me. If my knowledge of audio amplifiers is basic, I should see basic information in that section. But if I am a very advanced user of audio amplifiers, which Amazon can gather from my previous transactions and clicks, then the page I am looking at should be merchandized with high-end audio equipment. This is going beyond collaborative filtering. This goes toward being more relevant and targeted. It is all about cutting through the noise. The one thing that is true for the past 10 years is that the [number of] messages consumers receive has more than quadrupled. How do you cut through that noise; how do you become more relevant? That is what we are doing. There are other algorithms we use for predicting the propensity, the lifetime value or the spending potential of a customer. If we get down to the algorithmic level, there are several methods we use; anything from the pre-processing side, anything from a nonlinear principal component analysis (PCA) type model to single evaluation composition. On the declassification side we use logistic regression, neural networks, and internal-based methods.