In an era where online shoppers face endless product catalogs yet still spend hours searching without finding the right item, a critical question emerges: can AI truly understand shopping intent and deliver recommendations that feel personal, relevant, and context-aware?
At Levi9, a recent R&D initiative explored exactly that. The project demonstrates how recommendation systems can move beyond surface-level similarity matching by combining behavioral signals with product attributes to deliver suggestions aligned with real shopping behavior and how such systems can be designed for real-world use.
The Challenge: Beyond “More Items Like This”
Traditional recommendation mechanisms often struggle to capture intent, context, and style preference. While they can identify similar items, they rarely understand why users engage with certain products.
The team’s goal was to improve this experience by building a recommendation engine that reflects how people actually shop. Instead of showing “more items like this,” the system interprets behavioral patterns and combines them with product characteristics to surface recommendations that feel contextually relevant, not just visually similar.
What Makes It Different: Practical Personalization at Scale
The solution focuses on practical personalization rather than theoretical complexity. It is modular by design and can be adapted to different aesthetics, audiences, catalog rules, and brand identities – an essential capability in fashion, where preferences vary widely across markets.
Key differentiators include:
- Behavioral understanding combined with product logic, using collaborative filtering alongside content-based techniques
- Privacy by design, with models learning exclusively from anonymized interaction data such as views and clicks, without storing personal identifiers or sensitive attributes
- Hybrid training and inference setup, balancing immediate relevance with long-term learning potential
- Context-aware recommendations driven by demonstrated engagement patterns rather than surface-level similarity
Beyond the technical aspects, the project demonstrates the team’s ability to frame business challenges as solvable AI problems and deliver solutions that work in real environments.
From Data to Discovery
Delivering this solution required close collaboration across a multidisciplinary team. It included:
- Data science and machine learning
- Data engineering (pipelines, processing, model hosting)
- Software engineering (Python, APIs, deployment)
- DevOps and cloud expertise (OpenShift AI, model serving)
- Fashion domain understanding and UX thinking
The result is not just a model, but a complete recommendation capability deployed on OpenShift AI – used both as a development environment for training models and as a production-ready platform for secure, scalable model serving.
The system was built using Python-based ML pipelines and leverages proven recommendation techniques such as collaborative filtering (SAR-style item similarity), content-based signals, behavioral embeddings, and ranking models. Together, these approaches ensure recommendations reflect both user preferences and product logic.
What We Learned
Building a production-oriented recommendation engine provided valuable insights across the full ML lifecycle. The team gained hands-on experience with model training, tuning, and deployment on OpenShift AI, as well as with operating models in cloud-native, Kubernetes-based environments.
The project also opened opportunities to explore different recommendation strategies and evaluate how such systems behave under real catalog constraints.
Opportunities for Growth
The modular architecture enables future extensions, including event-driven approaches such as click and view tracking via streaming platforms like Kafka.
The deployment patterns and operational knowledge gained through this project are also transferable to Kubernetes-based setups and MLOps frameworks such as Kubeflow, reinforcing a cloud-agnostic approach to AI delivery.
The Path Forward
This project represents more than an R&D exercise. It reflects a methodical approach to building recommendation systems that balance technical sophistication with practical constraints, prioritize privacy, and support continuous improvement.
As personalization becomes a key differentiator in digital commerce, Levi9’s experience with hybrid models, production-grade ML infrastructure, and responsible AI design positions the company as a trusted partner for organizations looking to turn AI into tangible business value.
***This article is part of the AI9 series, where we walk the talk on AI innovation.***
In this article:
Levi9 Serbia
