Ozon Tech has made the recordings of 60 talks from the E-CODE conference, which took place in Moscow on September 12–13 and was timed to coincide with Programmer's Day, available for free. The two main stories of the ML/DS track are the migration of the entire marketplace recommendation pipeline to neural network models and the industrial measurement of product dimensions on a conveyor using computer vision. The recordings can be watched for free on the ecode.ozon.tech website.
What happened
Ozon Tech held its annual E-CODE conference in Moscow on September 12–13 and published recordings of 60 talks on the official website ecode.ozon.tech, while the first selection of the best presentations was released in the official Ozon Tech channel. The ML/DS track examined the evolution of Ozon's recommendation system: the pipeline stages from candidate selection to final ranking are gradually being migrated to neural network models with multi-task optimization for sales, advertising, engagement, and LTV, and the Large User Model architecture combines user signals, including negative feedback. A separate talk was dedicated to the industrial measurement of product dimensions on a conveyor, the belt of which moves at a speed of up to 1 m/s: a stereo camera, segmentation, and 3D reconstruction work in tandem, and a dataset of over 100,000 images was labeled semi-automatically using Grounding DINO with automatic object detection.
Context
E-CODE annually serves as a platform where the Ozon team showcases the inner workings of its ML systems, and the accumulated talks form a chronicle of how recommendation systems and industrial computer vision at a large marketplace are developing. Multi-task optimization of recommendations for several business goals is an established industrial pattern, so the novelty of what was presented lies not in the idea itself, but in the sequential migration of all stages of a large platform's pipeline to neural network models. The move toward a single user model, where behavior signals converge, resonates with the general trend toward unified representations of users and products. In the warehouse case, the key fork in the road is the trade-off between measurement accuracy, conveyor throughput, and solution cost: the team compared the ML approach with classical methods and point clouds in terms of quality, cost, and performance, and used open foundational models like Grounding DINO and SAM2 as a data annotation tool rather than as a final production component.
Why this matters for the industry
Ozon publicly demonstrated a working example of migrating all stages of a large marketplace's recommendation pipeline to neural network models that are optimized for multiple business goals simultaneously, rather than a set of disparate models. The Large User Model with signal aggregation, including negative feedback, is a practical step toward a "single large user model," and if this line continues, such an architecture could become the default for large e-commerce platforms. The CV team is shown a technique for reducing data preparation costs: semi-automatic labeling of a dataset of over 100,000 images using Grounding DINO and SAM2 is feasible with open-source today, and the "stereo camera + foundational models for labeling + 3D reconstruction" scheme looks like a typical template for industrial vision in warehouses. For startups, this is a signal of double cost reduction: data for industrial CV is getting cheaper due to semi-automatic labeling, and classical components of recommendation systems are being commoditized against the backdrop of the migration of pipelines to neural network models, so defense will have to be through the depth of integration with data and product.
Why this matters for users
The main thing for the reader is free access: recordings of 60 talks are available on ecode.ozon.tech, and without participating in the conference, you can figure out how recommendation systems, industrial computer vision, Wi-Fi geolocation, and supercomputers for ML at the scale of one of the largest marketplaces work. This is ready-made material for self-study and for checking your own architectures against industry practice. For those planning a career in ML, it is useful to know that the E-CUP competition this season was for students: the prize pool was 7.2 million rubles, and the tasks were set on real Ozon data.
What is still unknown / limitations
A conference talk without published metrics, ablations, and online experiment design is a demonstration, not a reproducible proof: the idea and workflow are transferable, but not a validated result. The sources do not report on ready-made tools for third-party teams: no code, API, or models have been made available. Assessments that a "single user model" and semi-automatic labeling using foundational models will become an industry standard are an interpretation of the trend, not a fact, and need to be verified through future publications and Ozon tech articles.
Sources
Author
Look at AI, editorial team