Building a New Internal Data Science Team

Problem

At my current company, I was hired initially to lead an effort to build an internal data science team from scratch.

The company was at a point where they had their original line of products, but, after gaining more market share, they felt that incorporating the data into what we were offering would help us up-level ourselves. We’re in the commercial real estate space, which comes with its own unique set of challenges. I think that, especially in the wake of the pandemic, people are realizing how useful consulting the data can be. It can provide a lot of context to your business.

At the time, we were not very data savvy. Part of my job was figuring out some way of presenting this information in a way that my executive team would be able to understand readily.

Actions taken

We spent a lot of time building up the foundation of the department. Our means of storing data at the time was not as elegant as it should have been. This was one of the first problems that I wanted to solve immediately.

The second thing involved upleveling our data science capabilities. There is ad hoc analysis, where you’re exploring the data freely and building a prototype model from what you find, discussing the results with the clients. To be a fully interactive experience, however, the data needs to continuously be integrated into the product.

Doing so requires any data science team to make predictions in real time, as opposed to on an asynchronous basis. Our data science workflow needed to be updated just as badly as our infrastructure. Helping my leadership and our clients understand this pipeline was another bridge that I really wanted to build.

We had a lot of initiatives going on, and I started to think more about the best way to scale our team around these changes being implemented. There were a couple of traits that we were looking for in terms of finding the right talent. We wanted somebody who was not afraid of wearing several hats, because we were still a relatively small organization. We wanted to find candidates who were just as scrappy at heart as we were.

Skills of analysis, literacy in the language of data, and experience in ML deployment were also attributes that we were looking for. Our ideal candidate needed to be able to contend with the full stack. We also wanted somebody with some experience in the real estate industry; this exposure would, in theory, help them relate the technical work to the business problems that we were trying to solve.

Real estate is an old industry; it hasn’t quite made the leap into the world of technology like a lot of other industries have. The data are numerous and very, very messy. I think that those of us in real estate technology are still trying to figure out ways of collecting cleaner data. After nine months of work, we were able to release our first working model. It is not perfect by any means, but we have been able to gain significant insights and to begin the process of iterating our predictive product with our clients.

Lessons learned

• The data can help you communicate what is possible and what is not to your leaders. A lot of this power depends on how ready the data is to present.
• We rely heavily on machine learning, which uses historical data to make prescriptive decisions. Really good data enables the system to make the best decisions possible. We did the legwork and collected a lot of good data over the course of several years, setting ourselves up for our current success now. What we are doing now would not have been possible without collecting clean data preemptively. Knowing things like this grounds my colleagues into the reality of what AI actually is and what it is capable of.
• In a larger company, you will usually be able to afford an expert in every domain. This was not the case; we were very specific when it came to the type of person that we were looking for. This allowed us to maximize every seat to be filled.