Growing a data analytics function – An evening with Balderton

1. When structuring your team, consider the ‘hub-and-spoke’ model

There are a number of ways to structure a data science team today, and many roles that can fit under the data science umbrella. So where should you start?For Chris and his team, the hub-and-spoke model has been invaluable. What this means is separating out your team into two core functions: the hub and the spokes. The hub is your central data team – a centre of excellence in charge of the platform, tooling and process standards. The ‘spokes’ are business domain teams, who own the data products for their domains.This structure gives essential balance. The spokes offer tailored insights from embedded experts. The hub offers honest assessment and evaluation from a neutral perspective, connecting the dots between otherwise disparate parts of the organisation. This is a key advantage, giving your team not only the ability to accurately measure but also to make competent, strategic decisions off the back of those insights.This is contrary to the ‘data-mesh’ approach. Data mesh is focused on a decentralised approach, allowing all users in an organisation to access and visualise insights from any data source, without the intervention from expert data teams.

2. As you scale, self-service analytics platforms are invaluable

Both Markus and Chris felt strongly that as you scale, it’s essential to have good self-service functions to democratise technical content.Markus highlighted the importance of teaching data literacy at Beauty Pie, in order to empower non-technical teams across the organisation to be able to understand and access insights independently. One way to do this could be by implementing a clear, accessible core dashboard to help people find the information they need and make strategic decisions easily, without being tethered to a member of the data analytics team.Data visualisation platforms like Looker can be helpful for this and provide huge value for startups, particularly in the early days. The idea is not to have thousands of dashboards, but a way to democratise insights clearly through a ‘semantic layer’, as Chris called it, that can bring value across the business. And for DS specifically, having a “hub” that is intentional about maintaining feature stores and model registries empowers the “spokes” to follow best practice when building ML models from a consistent source of truth.

3. Building the perfect tech stack: dbt all the way

For Series A-B+ startups that have reached sufficient scale, our panellists recommended dbt as a must-have. It is affordable and straightforward to implement, enabling running tests more easily. With dbt this leaves room for building new features while being safe in the knowledge that changes in data quality are flagged by CI checks.Intermediary platforms can also help data science teams that might be using Jupyter Notebooks, allowing Jupyter Notebooks to reference dbt projects. Markus recommends keeping dbt in SQL, as well as using the semantic layer that dbt offers.

Thank you to our panellists for coming along and sharing their insights. If you’d like to be kept updated on future events or conversations on this topic please email [email protected].

Related content