// the find
decodingai-magazine/personalized-recommender-course
👕 Open-source course on architecting, building and deploying a real-time personalized recommender for H&M fashion articles.
A five-module course teaching you to build a real-time personalized recommender system using the H&M Kaggle dataset, two-tower neural networks, and the 4-stage retrieval-ranking-filtering-reranking pipeline. It's aimed at ML engineers who want to see how Netflix/TikTok-style recommenders actually get wired up end-to-end, not just trained in a notebook. The live demo and GitHub Actions deployment make it more than a toy.
The FTI (Feature/Training/Inference) pipeline separation is well-enforced in the code structure — features, training, and serving are genuinely distinct modules, not one giant script. Using Polars for feature engineering is a solid choice over pandas for this scale of data. The two-tower model implementation is clean and the course explains why retrieval and ranking are separate stages rather than treating it as one black box. Cost is realistic: modules 1-4 are genuinely free, and the ~$1-2 OpenAI cost for module 5 is honest rather than buried.
Hard dependency on Hopsworks for the feature store and KServe for deployment means you're learning Hopsworks as much as recommender systems — if you want to adapt this to AWS SageMaker Feature Store or Vertex AI, you're mostly on your own. The notebooks are the primary learning artifact rather than the recsys Python module, so the "production code" framing is aspirational; real production systems don't run training via Jupyter. Module 5 uses OpenAI for LLM reranking with no mention of alternatives, which is an odd choice in a course that otherwise stays free. The dataset is 31M transactions from 2018-2020, so anything you learn about temporal dynamics or cold-start handling won't reflect how H&M's actual catalogue behaves today.