Case Study · Retail
Defacto DSO: four years from the founding of a data science office
We started in 2021 as mentors at the hiring Datathon; it continued to the end of 2024 with the team's training, the DSO's database and server infrastructure, and more than thirteen projects written and put into production. Today the DSO is a directorate serving Defacto's retail analytics, AI and business intelligence needs.
- 4 years
- Duration
- 13+
- Projects
- Stevie 2023
- Award
- SQL · Python · data science
- Training
From founding to handover documents
AI, algorithms and business intelligence
IofTheCustomer — customer review analysis
Onboarding programme after hiring
Context
In early 2021 Defacto was founding its Data Science Office (DSO). There was no team yet; hiring would be done through a Datathon. Lumtify joined that Datathon as mentor — preparing the questions, evaluating, then the hiring itself. Work began in March 2021 with a team of 17 + 6; the team was completed with the hiring of a director.
The scope of the engagement stayed the same for four years: generating project ideas, mentoring the team, coding projects and putting them into production, building the infrastructure, training to raise the team’s technical competence, defining reporting needs. In 2023 the Lumtify team grew to four; 2023 and 2024 were completed as “we” rather than “I”.
Infrastructure first
On the DSO’s first day there was no database, no virtual server and no permissions. If a team was going to be built, it first needed somewhere to work. A roadmap was drawn with the technology team and built in order:
- DSODB — the database where all of the DSO’s procedures, tables and views live, designed with a notation similar to the company’s data warehouse. Which table goes in which schema and how it is named was documented and passed on through training.
- Three virtual servers — the machines where Python projects run, where SSIS packages are scheduled as jobs on their own SQL Server, and where the Power BI gateway sits. Because the DSO had no permission to turn a procedure directly into a job, it built its own scheduling structure through SSIS packages.
- Log structure — from day one every SQL and Python procedure writes to the same log table: start time, time between steps, the step that failed. Every registered project is monitored from here.
- Product development process — table and column naming, procedure and view rules, delta tables, a periodic maintenance list (server restarts, disk, TempDB, job failures).
This layer is what made projects person-independent: if a project runs as a job on a virtual server, it keeps running when the person who wrote it is on leave. Setup, permissions and maintenance were carried by Lumtify for four years and handed over, documented, at the end of 2024.
Growing the team
An onboarding programme was prepared for the team hired through the Datathon: SQL, Python and data science training (by Lumtify), case studies, assignments and starter projects for adapting to the sector, then the forming of teams and the first real projects. Training was not a one-off but a mentoring relationship spread across all four years: on some projects we wrote the code, on others we only set the direction.
In the same period we supported Defacto’s data-driven transformation: each department’s data maturity was assessed, year-end targets were set, and data structures, reports and analytical solutions were planned accordingly.
Projects
Our habit of naming projects paid off here: a name made a project easier to own and easier to turn into a brand inside the company. The projects we were part of over four years, and our role:
| Project | What it does | Lumtify’s role | Status |
|---|---|---|---|
| IofTheCustomer | Classifies product reviews by sentiment and category, routes them to the responsible team | AI + project management | Live for 3 years · Stevie 2023 |
| Sorting | Product ordering on e-commerce category pages | Coding, integration, maintenance | Live |
| ELOSS | Lost sales from stock-outs on e-commerce and marketplaces | Coding, integration, maintenance | Live, weekly |
| NoSales | Stock unsold for 28 days in stores and e-commerce | Coding, integration, maintenance | Live, weekly |
| BSA | Highlighting city-level best sellers in stores | Coding, integration, maintenance | Live |
| SearchTerm | Search box query analysis and spelling correction | Mentoring | Live |
| PAT | Meeting productivity dashboard from Teams and Meet data | Data infrastructure, procedures, dashboard | Live |
| JUMP-UP | Weekly KPI tracking for top management, year-on-year | Data infrastructure, procedures, dashboard | Live |
| DISCO | Sales forecast 14 days after a price change, markdown scenarios | Model, daily pipeline, demo interface | MVP; never went live |
| preDefaction | 6-week sales forecast per store | Data preparation, model development | Shelved |
| Basket Analysis | Products bought together, cross-sell | Mentoring, end-to-end integration | Live |
| Capacity | Checking store capacity split ratios against actuals | Technology integration | Live |
| Lot Optimization | Re-boxing season-end returns with assortment in mind | Mentoring | Built by the DSO |
JUMP-UP deserves a page of its own but its technical detail sits on the BI side: it shows Defacto’s critical KPIs — revenue, units, gross margin, customer count, AOV, cover, GMROI, DIO, broken size runs, OTD/ALT, capacity utilisation — on a single nine-page Power BI dashboard, weekly and compared with last year. Data is prepared every Monday morning by scheduled queries in BigQuery and procedures in DSODB; pages from the NoSales, DIO, Residue and Capacity dashboards are reused.
BSA — the Best Seller Algorithm analyses e-commerce and marketplace sales by city and determines the products to highlight in each city’s stores. “Top seller” alone is not enough: units, margin and sales velocity are weighed together, and the list does not reach a store without passing the checks of whether the product is available there and whether there is stock to display. A bridge turning online demand into store display; live as a scheduled job.
SearchTerm analyses the queries in defacto.com’s search box: it corrects spelling (“pantlon” → trousers) and measures the chain starting from each term — views, basket, purchase, conversion, revenue — end to end. Behavioural data from Google Analytics and Defacto’s product and sales data merge in one analysis. The DSO team developed it; we mentored.
preDefaction forecasts the next six weeks of sales for each store from sales history, store size, city, region and look-alike matches for new stores. Pandemic closures and restricted hours were labelled at store–week level to repair the data; several machine-learning and time-series models were raced and the best per store chosen with BestFit. The model produced good results, but no decision point was defined where the forecast would be a direct input, so it was shelved — the reason we now ask “who will use the forecast, in which decision” at the start of every forecasting project.
Basket Analysis derives products bought together from receipt and basket data at option, class and buyer-group level, producing both cross-sell suggestions and marketing insight. Capacity compares the ratios product management uses to split store capacity into sub-hierarchies against actual sales, margin and revenue, and suggests ratios. Lot Optimization is an iterative optimization that re-boxes broken-assortment products in the returns warehouse for outlet and markdown shipments.
Support work
To establish the DSO as a solution centre within the company, requests from other departments were also met: moving e-commerce payment analysis from spreadsheets into the database and onto a dashboard; loading third-party competitor crawl data into DSODB and building comparison dashboards against Defacto’s data; centralising overseas marketplace data that did not arrive through integration; measuring the adoption rate of the in-season management software’s outputs.
Outcome
At the end of four years the DSO had become a directorate answering all of Defacto’s retail analytics, AI and business intelligence needs. Projects run independently of individuals; the infrastructure, job list, log structure and maintenance procedures were left with the team in handover documents.
For us Defacto was as much a school as a client: the chance to see end to end how a data science office is built from zero, how a team grows and how projects are kept alive. The “infrastructure first, then visibility, then the decision” order we applied at Derimod and Kalekim was learned here.