E-commerce
Weekly spreadsheet reports replaced by a live dashboard and a demand-forecasting model, with product recommendations served from the same event stream.
The commercial team ran the business on reports exported from the production database every Monday. Queries on the order tables took minutes and occasionally slowed checkout, so exports were limited to one run a week.
The client also wanted personalised recommendations on product pages, but had no event pipeline, no feature store and no one on staff who had put a model into production.
How things ran when we started, and once the work shipped.
Approach
4 phases over 5 months, with a working demo at the end of every week.
Weeks 1–5
Defined an event schema with the product team, instrumented web and app clients, and streamed events through Kinesis into S3 and Snowflake.
Weeks 4–10
Built dbt models for orders, inventory and customer cohorts, and a Next.js dashboard on top with role-based access for finance, buying and marketing.
Weeks 9–18
Trained a gradient-boosted demand forecast per SKU and a two-tower recommendation model, then A/B tested recommendations against the existing bestseller list.
Weeks 18–22
Added weekly retraining, drift alerts and a model registry, and paired with the two data analysts the client hired so they could own it.
Production databases are no longer queried for reporting. Everything flows from a single event stream into the warehouse, and models read features from the same source.
“Our buyers used to wait until Monday to see what sold last week. Now they reorder from a forecast every morning, and the recommendations paid for the whole project within the first quarter.”
Tell us where things stand today. On a 30-minute call an engineer will sketch how we would approach it, the timeline and a price range.