This open-source course teaches how to architect, build, and deploy a real-time personalized recommender system for H&M fashion articles. It is targeted at ML/AI engineers, data engineers, data scientists, and software engineers interested in production-grade recommender systems and MLOps. The course provides a hands-on, end-to-end implementation guide using modern tools and best practices, enabling users to create scalable and modular ML systems.

How It Works

The course follows a 5-module structure, covering system architecture, feature engineering with Polars, neural network model training (two-tower network), and deployment on Kubernetes via Hopsworks Serverless and KServe. It emphasizes an FTI (Feature/Training/Inference) architecture and a 4-stage recommender design, incorporating a feature store and model registry for MLOps. LLM techniques are also introduced for enhancing recommendations.

Quick Start & Requirements

• Install/Run: Follow the INSTALL_AND_USAGE guide. Notebooks can be run locally or on Google Colab.
• Prerequisites: Basic Python and ML knowledge. No GPU required; Google Colab or GitHub Actions are supported for compute.
• Cost: Free, except for optional OpenAI API usage (~$1-2) in Module 5.
• Links: Live Demo, INSTALL_AND_USAGE

Highlighted Details

• Hands-on implementation of a real-time personalized recommender for H&M fashion articles.
• Utilizes the Hopsworks AI Lakehouse for MLOps practices (feature store, model registry).
• Employs a two-tower network for user and item embeddings and a vector database for retrieval.
• Includes a Streamlit web interface for demonstration.

Maintenance & Community

• Collaboration between Decoding ML and Hopsworks.
• Contributors include AI/ML engineers from Decoding ML and Hopsworks's Engineering Team.
• Questions and troubleshooting can be directed to GitHub Issues or Hopsworks's Slack.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive for educational and personal projects, requiring attribution.

Limitations & Caveats

The live demo may experience a 1-2 minute warm-up time from a zero-instance scale. Module 5's LLM enhancements incur minor OpenAI API costs. The course focuses on engineering and implementation rather than theoretical model optimization.