Modulai developed and deployed a custom-built recommender system for Lindex, one of the largest retailers in the Nordics. The system processes data on approximately 60,000 products and purchases from over 2 million customers, using advanced NLP and image processing alongside collaborative filtering to deliver personalized, multi-channel recommendations.
Outcome
+35 MSEK
Yearly additional revenue
+2–4%
Gross revenue increase
2M+
Customers served
Challenge
Lindex is a major retailer active in the Nordics and throughout Europe, with several million registered customers and a prominent presence in women's wear, lingerie, and kidswear. Customers expect retailers to provide recommendations based on their browsing and purchasing behavior — a feature that both generates sales and provides a sense of being cared for.
Solution
Customers were characterized from their historical transactions, while products were described from multiple angles: category hierarchies, attributes such as color, texture, and material, product text, and product images. By combining what products are with how customers behave, the system produces recommendations at the individual level, drawing on both content and collaborative signals rather than purchase history alone.
Tools
Product text was processed with classical NLP techniques alongside Universal Sentence Encoders, FastText, and BERT, while product images were turned into dense representations using ResNet and convolutional autoencoders. Recommendations were generated using matrix factorization and gradient boosting models. Preprocessing, training, validation, and inference pipelines were built in Python on Azure, with Snowflake, Databricks, and Spark orchestrated through Apache Airflow, and text feature extractors developed in TensorFlow.
Value created
The recommender system was deployed into production, delivering individual-level recommendations across channels to a customer base of over two million. By combining purchase history with what the products actually look like and how they are described, it generates recommendations grounded in both behavior and product content, rather than purchase patterns alone, at the full scale of Lindex's catalog.
Learn more
