I have been with Poshmark for four years and work as a Senior Director of Data Engineering. I am responsible for managing the company’s data infrastructure, machine learning and partner integrations. Before Poshmark, I worked with Digisight Technologies, a medical startup, where we developed an algorithm for a managed cure for diabetic retinopathy. I am not a data scientist, but I manage machine learning teams. These are full-stack teams responsible for analysis, building data science models, and for productionizing those models. What drew me to the area was my interest in business intelligence. Then, working on big data systems at Poshmark. This led me into further research into what it takes to develop a model, based on the data we had at Poshmark. So, you could say that industry interests inspired me to do data science in order to develop models that help the business.

When I first came in touch with the field, the focus was primarily on data warehouse solutions and understanding users and user behavior through transactional data. We have since seen a lot of growth in the behavioral data collected by companies: how much time users spend on a platform, what interactions they perform, what products they are interested in, how they can be kept engaged and so on. All of these questions can be answered using machine learning.

A little over a decade ago, companies like Netflix were creating recommendation challenges to draw people in to learn about the field. There has since been a big shift in terms of algorithms and how they are being adapted. Now, we’ve reached a point where machine learning is nearly at the point of being commoditized. Earlier, you needed team members with a deep understanding of the models. Now, the models themselves are being developed as a product. The industry is evolving such that you don’t need a deep understanding of models, but can develop them so long as you have the ability to play with the data.

The lesson I have learnt in this journey is that you have to pick the right problem and make sure you have the right dataset for it. One must ensure that one has good quality data that one can rely on to develop a model. But not everything is driven by a model, not every problem needs a data science solution. So, it’s important to understand for each problem how far you must go to solve it.

The technology landscape is changing so drastically that it is always challenging to keep up. And there are new trends to understand out there in the market and bring into the organization for its benefit. That’s very exciting for me. Given that our business is growing, I always have to work on the most current tools and understand how to scale them. And if a tool doesn’t scale, I have to figure out what I need to do to replace that tool. The industry is changing so much, it gives me joy to continuously learn and challenge myself.

Identifying and understanding data seems to be a large challenge. There are some team members who would love to do just that—sit with data, understand it, and spend hours and hours exploring data. And there are others who, once someone else develops the models, are more interested in productionizing them. What I’ve noticed is the different backgrounds of team members that come into this profession. And as a leader you have to be able to work with different kinds of team members and skill sets. Some team members are so intimately familiar with data and algorithms but have challenges explaining their models and their benefits to business users. That involves communicating in simple language that everyone can understand. So, you get to play that part of translating from one group to another. There are many facets to making a data scientist effective and finding the right mix of skills seems to be a challenge. And, in general, this is a field of high demand with a lot of competition, and hiring is extremely difficult for these roles.

Many of the graduates wanting to do data science that I have seen come in with the sole intention of developing models. But before you develop the models, you have to do all the groundwork necessary to get there. And some people may not be interested in doing that. Not establishing that requirement is one mistake companies make.

Further, with any hire, you want to make sure that they are culturally fit for your organization. For instance, some organizations may be engineering-driven or have a data-science-first methodology. Candidates with a similar predominant focus would be considered as the best fit. Some candidates come from a research background with much longer time frames for projects, while what we call practical data requires models to be developed and put into production within typical time frames of six months to one year. So, you want to see candidates show qualities that fit those requirements. Finally, for a startup like us, we need to know if prospective team members can wear many different hats and do everything needed to make their project a success.

When you bring a new team member onboard, you’ve already spent significant effort and time in hiring them. So, you want to make sure that they are successful. Therefore, it’s important that you properly vet them during the interview process to know whether they are going to fit culturally or not.

We have tried to do this in both ways. Initially, we began with one of the best collection frameworks for high quality data. To use that data, we developed what we call data applications, which aren’t actual models but provide some insights into the data.

At the second level, we looked at what models we could develop on the strength of this data. This was somewhat driven by engineering, by the desire to design and develop models. There has also been the demand from business, from growth and marketing teams, for the data science needed to run efficient campaigns.

Further, engineering also developed a recommendations platform which has been in production for a year, and which has begun delivering value to the company.

We are fairly open to team members developing or bringing new tools into our ecosystem. Right now, we use Spark ML, Pi Spark, and Scikit-learn, among others. We also use SageMaker to a certain extent in research. SageMaker also comes with many different packages and a variety of other tools embedded in it. At Poshmark, we have two data teams: the first is an analytics and machine learning team, while the second is the data engineering team that I lead. My team comprises of data engineers and machine learning engineers. Our machine learning engineers are full stack data scientists and software engineers.

We receive applications from candidates from a variety of backgrounds. So, we first review resumes for technical fit, based on their backgrounds and the work they have done in the world of data science and machine learning. Those candidates that we feel are a potential fit receive a take-home assignment involving a business problem to be solved using a modelling algorithm. This is followed by a phone screen with a software engineering or machine learning algorithmic focus, depending on the candidate’s strengths.

Then, in the new normal, we follow up with virtual onsites. These involve four to five hours of interviews with the head of the data science team, a machine learning engineer, a hiring manager, and a big data engineer. For machine learning engineers the focus is on their ability to design, architect, scale and integrate a model – basically the overall lifecycle. Finally, we also have a cultural fit interview with a different panel.

Looking at a resume, we want to get a sense of whether candidates list projects associated with a business or ones they explored as part of learning how to model data. Next, we look at whether the resume gives a sense of the details of problems solved, algorithms used, and so on. Further, while we don’t expect our candidates or potential hires to have the same tools as what we use in the organization, we do check if they have something comparable. Finally, we look for signs of whether they can scale up, whether they have developed models in a distributed environment or have only created models on a single stack. Those are the things I look for in a candidate with experience.

With fresh graduates we look for potential to learn. We look for whether candidates have done a little bit of work to align their applications with the job requirement, are able to grasp what the position describes, and highlight what they have learned. We are generally very flexible with resumes because our take-home assignments are appropriately difficult, and we are confident of interviewing candidates who are able to put in the effort and come through that stage.

In interviews, we look for an ability to deep-dive on past projects. We generally work with whatever candidates know – we ask them to explain their projects; the challenges they faced; how well they know their business problems, data sets, and algorithmic work; and how they plan on revising their models.

We are also big on cultural fit. We expect individuals to show us that they are not afraid to fail. Also, they should be willing to work with different organizations because, as machine learning engineers, they have to work with product teams, data engineering teams, users to get user feedback, and so on. We generally look for people who have a can-do attitude and are willing to do whatever it takes.

We put a lot of thought into creating job descriptions and our job descriptions will include six-month and one-year accomplishments. When candidates review, understand and internalize the accomplishments listed, the interview conversation will generally be very organic. I would say, spend time on understanding the job description and what the company does and speak to that. We generally look for

• Software engineers that can work with algorithms
• Overall experience with building and deploying models (not simply deploy only)
• Ability to monitor and improve model performance
• Ability to understand the business problem and communicate with business users
• Ability to learn and take initiative
• Culture and team fit

We are open to training candidates who bring in strong skills in one area but may be less skilled in other areas. We are hiring not only to solve current challenges, but also future challenges that we have not yet discovered.

I personally think they should learn and understand different toolsets, as well as try to play with a variety of sample projects and datasets available on public sites to continue to improve their skills. Some of the Big Data technologies & Machine Learning frameworks we are typically involved with include SparkML, TensorFlow, PyTorch, and Keras.I would say that they should become familiar with all open source tools and networks consistently.

To stand out from the crowd, candidates should do some research into the organizations that they are applying to and see what kind of tool sets they deploy. One practical way of doing this is to log onto job sites and compile the various job descriptions to highlight the most common tools, and to familiarize oneself with them. When applying for jobs, candidates should also highlight what they can bring to the table. They can do so by finding ways to understand what the organization is developing.

Our mission is to make shopping simple and fun by connecting people around a shared love of fashion, while empowering entrepreneurs to become the next generation of retailers. We are a social-commerce company. We face challenges such as ecommerce and social networks and have to solve them simultaneously. We have a wealth of interesting, high-quality data waiting to be utilized for deriving value using models and applications. We are currently solving several interesting data science problems involving recommendations, feed ranking, search optimization and fraud.