Database Reference
In-Depth Information
I mean all of it—what they told them over the phone, which sales scripts, if
any, were used, what they told them, how many items they bought, and so
forth. Basically, anything that has been input into the CRM is useful for us.
Once we have this data, it is passed to the web crawlers we built. These crawl-
ers crawl the web and get as much information possible about the persons
being sold to and the companies that those people work in. For the people,
we want to get a better picture of who they are and how they're represented
on the web. For example, it would be helpful to know if they are a technical
or nontechnical people.
For the companies, we want to get a better picture of who they are—not only
how they're represented on the web but also the more relevant financial and
operating information. For this analysis, we crawl their web pages, we do a tax
analysis on them, we categorize what the company specializes in, and we buy
traffic data about a company, so we have data on how many people are going
to the potential customer's website.
We also get access to our customer's website data and how their potential
customers have interacted with it. For example, did a person download their
white paper, what pages did they view, and so on. We have a large amount of
behavioral data about a potential sales interaction before the salesperson is in
a conversation with his or her prospect.
Then, after having aggregated data on the prospect, the prospect's company,
and behavioral data from our customer's website, we go to the math, machine
learning algorithms, and models. We want two things from this data. The first
is to understand deeply how sales transactions have happened in the past. The
second is to be able to calculate, based on the historical data we've seen in the
past, the probability of a new person becoming a customer and the actions we
should take to make them a real customer.
On our side, we build these models automatically. This includes adding in the
data. The entire process from the customer's point of view is done in less than
ten minutes. Everything else is done completely automatically on our servers,
like the crunching of the data, gathering of the data, feature engineering, and
so forth, and it poses a lot of challenges. So that is a simple description of the
scheme of how SalesPredict works.
Gutierrez:
What challenges have you faced as you've grown the customer
base, technology, and data?
Radinsky:
The challenges in these types of models emerge from moving from
a simple scheme to the real world. First of all, our market is very demanding.
Our customer's historical data and distributions change all the time—their
perfect customers today may not be who it was, say, two months ago. As
everything changes, this model has to be rebuilt automatically as the data
changes—approximately every week. More than that, we have to supply
