Seneca Widvey is a data science leader with over a decade of experience in AI, machine learning and analytics. As Director of Data Science at Audacy, he drives innovation in recommendation systems, forecasting models and MLOps. With expertise in applied mathematics and AI, Widvey excels in building scalable data solutions and leading high-performing teams to deliver business impact.

Through this article, Widvey highlights key principles for building scalable data science solutions, particularly within a corporate media environment to effectively navigate technical challenges and deliver impactful, business-aligned results.

Building scalable data science music and radio solutions requires navigating a sea of technical complexity. Much like sailors guided by the North Star, enterprise data science teams can rely on key principles to steer through these challenging waters. What I’ve learned after five years at a media company undergoing transformation is that there are three key principles for navigating our field in a corporate environment; ensuring data quality, addressing stakeholder needs and favoring practical simplicity over complexity. These simple principles allow for both technical rigor and business impact. 

A common expression in the data field is “Garbage in, garbage out.” While this phrase highlights the importance of quality data, it lacks the nuance needed to think of enterprise data as a product. Quality data is multidimensional, encompassing accuracy, availability, consistency and timeliness, to name just a few. Almost like the four directions on the sailor’s compass, these dimensions are critical to building scalable and impactful machine-learning solutions.

● Accuracy: If data values do not reflect actual user behavior, recommendation systems will be inconsistent and fail to meet user expectations.

● Availability: Unavailable data sources can disrupt production model workflows. Missing or inaccessible data can halt updates, leaving users with stale or outdated suggestions for content exploration.

● Consistency: Inconsistent data, such as column values becoming NULL or missing, can undermine model training. This could result in skewed predictions or even the inability to generate recommendations altogether.

● Timeliness: A good recommendation system needs fresh data. Without timely updates, recommendations quickly become stale, leading to a poor user experience.

To create and scale a machine learning recommendation system, a data science team must collaborate closely with their data teams to treat data as a product. This "data-as-a-product" mindset ensures that data is curated and maintained with the same thoroughness as the recommendation system itself. Product thinking applied to datasets is essential to delivering impactful recommendations. 

Getting from data to design requires a deep understanding of stakeholder needs and the business impact of your work. In other words, you need a map that data science and business stakeholders can agree on to achieve your scalable goals. A KPI holds no value unless you understand the problem it addresses and its broader implications. To create your data product map, you must align. And, alignment requires asking questions—lots of probing questions. 

“Building data science capabilities requires that you have your business North Star, a compass and a map. Once you have those as outlined above, navigating complexity becomes smooth sailing.”

Early in my career, I worked on a fraud detection algorithm for stakeholders and I made the mistake of taking their initial requests at face value. I didn’t ask the critical questions is ‘How will this be implemented? How often does it need to update?’ These simple yet essential questions would have saved significant headaches when trying to integrate the algorithm with our backend engineering systems. This experience taught me the importance of putting myself in stakeholders’ shoes to understand their priorities and how they intend to use the solutions we build.

A great example at Audacy is A/B testing for click-through rates (CTR). While optimizing CTR seems straightforward, focusing solely on it can overlook other important metrics. For instance, a feature might boost CTR but reduce listening hours, leading to lower ad revenue and a net negative impact. By considering the broader business context, we can balance multiple metrics to ensure alignment with strategic goals and long-term success.

Asking the right questions and stepping back to see the bigger picture helps bridge the gap between technical design and business outcomes. This approach ensures that data science efforts not only solve the right problems but also deliver meaningful and measurable impact.

As we understand the business problem, the next question becomes ‘How do we scale and create effective data science models?’ Often, in the excitement of building complex deep-learning or generative AI models, we overlook the power of simplicity. In many cases, a simple design can be more effective than state-of-the-art, sophisticated models.

Simplicity offers key advantages like faster development, easier understanding and immediate feedback. For our data science team, creating forecasting tools often starts with simple models like moving averages or ARIMA. These accelerate development, establish a baseline and guide improvements. Simplicity also benefits system design, enabling the gradual addition of features like monitoring and error handling. Once the system is stable, complexity can be introduced step by step, ensuring each addition is purposeful.

Navigating the complexity of data science solutions does not have to be troubled waters. A thoughtful balance of quality data, stakeholder collaboration and simplicity in design will help bridge the gap in delivering technical execution and business impact. To belabor the anecdote, building data science capabilities requires that you have your business North Star, a compass and a map. Once you have those as outlined above, navigating complexity becomes smooth sailing.