What we built, what we measured, and where the limits are.
From siloed systems to one source of truth, then predictive models and a marketing mix model
One source of truth, then models the brand’s own team now runs.
Challenge
The brand was growing fast, but departments ran on separate systems with different KPIs and conflicting definitions. Leadership had no single, trusted view to steer spend or forecast revenue.
What we did
A modern data stack on AWS, dbt and Looker unified separate sources into one governed source of truth. We then built a revenue forecast by customer cohort and a marketing mix model designed to be retrained. As the brand built its own team, we trained it and handed it over.
Result
Every team works from the same governed numbers, with sharply less reporting overhead. Applied to a real March window, the optimal mix predicted lifts in core revenue of 12% in Canada and 18% in the US, about US$4M (~5%) in predicted annual revenue.
How we measured it
Marketing mix model fitted on historical spend and revenue, then applied to a real March window. Lifts are modelled, not observed.
Source: Client data and DS analysis, predicted
Adding visual browsing and smarter recommendations to an AI shopping assistant
Production AI microservices that help shoppers find what they want, fast.
Challenge
The platform was rebuilding its AI shopping assistant around a large language model. It needed visual browsing, and recommendations tuned to drive sales instead of weighting every product tag equally, fast enough for live chat.
What we did
A visual browsing microservice that reads the chat and the catalogue, estimates what the shopper wants, and picks images that narrow preferences while showing range. Plus a recommender model that weights product attributes by their effect on sales, with cold-start baselines for new retail sectors. Both built as documented APIs.
Result
Both microservices met the platform’s acceptance criteria and run inside its own pipeline, returning representative products in under 10 milliseconds for catalogues of up to 10,000 items.
How we measured it
Visual browsing was judged on information gained per question, how quickly options narrowed, and response time; recommendations on engagement, conversion and revenue.
Source: DS measurements
What people say about the work
DS really helped our team organize our retail and CRM data into value clusters and augment customer value segments. On top of the delivery of the models, the knowledge transfer to our team was invaluable.
Being data-driven takes more than having data. Data is only the first step. At DS, we use it to describe what happened, support decisions and improve outcomes.
Want results you can stand behind?
Tell us where you are today. We’ll tell you where we’d start.