I was talking about our data infrastructure and use cases of machine learning that have helped us to be the leader in the recipe kit market in terms of the proposition we offer to our customers. Our customers can choose from 22 recipes and get their boxes delivered in 2-3 days from a choose 6 delivery days a week (we don’t deliver on Fridays at the moment).
Snowplow plays an integral part allowing us to collect very granular data on our customers’ interactions. Besides collecting interaction events from our website, mobile apps and emails, we also use Snowplow as a unified log collecting backend events, which completes a single customer view.
On top of Snowplow and other data sources, e.g. transactional database, warehouse management system, CRM etc., we use Airflow to orchestrate ETL and other tasks to model the data. To visualise and explore the data we use a combination of Periscope Data, also heavily used for reporting, and jupyter notebooks.
In the second part, I was talking about our use cases of leveraging the (Snowplow) data. We do various sorts of data-science at Gousto, from analytics to predictive modelling and optimisation. For the sake of time, I focused on two very exciting use cases:
- churn prediction and
- automated menu design.
I will write more about those two projects in the future, but for now in a few sentences.
The churn prediction helps us to find customers more likely not to place an order in the near future, so we can act in order to retain them. The idea is to use granular events data and machine learning to estimate for each customer how likely is to place an order in the near future. The main challenges are to define churn and find data (events) with predictive power. But equally, it’s also very important to find the right and efficient actions to take in order to make it profitable.
As a recipe kit service, our menus are an integral part of the business, therefore it’s important to make them as good as possible. But, the million dollar question is what are good menus. To answer that question we need to understand our recipes first and therefore we built a recipe ontology and an internal Slack bot to collect and understand recipe similarity data. On top of that we use Snowplow data to understand customers’ tastes so we can provide as many popular menus as possible.
It was really fun presenting at Snowplow Meetup. We got some great feedback and more exciting ideas how to use and leverage Snowplow.
For all those who missed my talk I embedded the slides below.
Head of Data Science