Scaling E-commerce Data Quality: Automating Product Image and Data with Google AI

Boozt @ Google AI Day 2025 in Copenhagen, Denmark

This blog explores how Boozt is using artificial intelligence on Google Cloud to automate and improve product image and data workflows. It highlights how high quality data and scalable AI architecture enable faster product launches, better search relevance, and more engaging customer experiences.

Authors: Jonny Johansson, Data System Director & Steffan Mølbæk Andersen, Data Intelligence Director

Boozt handles a vast number of products each year, which calls for robust and scalable technical solutions to manage both product data and imagery. With more than one million product photos processed annually from a wide range of suppliers, ensuring consistent quality and presentation across the site is a complex and ongoing challenge.

Our engineering focus is on leveraging AI to transition from a resource intensive, manual image and data approval process to a highly efficient, semi automated pipeline built on Google Cloud Platform (GCP). This project aims to deliver tangible business benefits: ensuring data consistency, improving on site search, and maximizing product discovery, all of which lead to higher conversion and faster time to market.

This work was recently presented at Google AI Day Denmark 2025 in Copenhagen, where we shared the architecture and results with a full audience.

Main stage at Google AI Day Denmark 2025 where Jonny Johansson, Data System Director & Steffan Mølbæk Andersen, Data Intelligence Director spoke.

Automating the Product Image Approval Workflow

The core challenge in image processing is standardizing output despite varied supplier inputs (lighting, composition, background). We are redesigning the approval flow to incorporate automated checks using Google Vertex AI together with standard models such as Gemini Flash 2.5 Pro, but also with custom trained models.

The system’s current phase intentionally combines automated prediction with human review to ensure quality and build confidence. Our roadmap is toward a fully automated process supported by human oversight for exceptions and continuous improvement.

Key AI Powered Quality Gates:

• AI Powered Attribute Prediction: We use Vertex AI models to predict key product attributes upon ingestion, often achieving correct predictions on the first pass, significantly reducing manual data entry. This includes predicting up to three levels of product category (e.g., Women > BOTTOMS > Trousers) and color.
• Image Position Check: The system verifies the correct labeling of product views (e.g., Main, Back, Model, Detail). This check is important for selecting the optimal “main” image to display, a decision that is category dependent (e.g., model shots for dresses versus contextual shots for home goods).
• AI Background Removal: To enforce a clean and consistent aesthetic, we utilize a custom Python Cloud Function running a pre trained PyTorch model to remove image backgrounds.
• Color Verification: Automated checks ensure consistency between the assigned product color and the visual color across multiple image sources will soon go live.

The direct outcome of this technical investment is a reduction in manual friction, shorter time to market for new items, and higher image quality, which is intrinsically linked to improving customer confidence and conversion rates.

Jonny Johansson, Data System Director at Boozt

Leveraging Product Data for Discovery and Search

High quality, accurate product data is the fundamental input for our platform’s core discovery features. We focus on two data driven use cases built on this enhanced product DNA.

1. Smarter Search

Our Smarter Search initiative focuses on enriching the product data feeds that power the search engine. By ensuring a reliable data foundation (clean images, robust structured data, and accurate attributes), the search model receives clean signals to learn from. This enhanced product content, generated through our AI pipeline, directly improves the ranking and relevance of search results for customers.

2. Similar Items

The Similar Items feature aims to expose product alternatives that align with an item’s “DNA” by learning relationships between customer interaction patterns and content attributes.

Architecture and Results:

This feature is built entirely within the Google Cloud Platform ecosystem:

• Data Backbone: BigQuery serves as the central data warehouse.
• AI Modeling: Vertex AI is utilized for the core similarity modeling.
• Orchestration: Management of distributed processing and event pipelines is handled by services including Cloud Composer, Dataproc, and Pub/Sub.

The results of this data architecture are highly encouraging: the launch of the new similar items model for our Home category delivered an almost 4 times increase in clicks to related products. This significant spike in engagement is clearly visible in our CTR data over time. Such outcome validates the approach, confirming that disciplined data quality combined with targeted AI services translates directly into strong discovery and higher customer engagement.

Steffan Mølbæk Andersen, Data Intelligence Director at Boozt

Takeaways and The Road Ahead

Our progress demonstrates a core principle: High quality data is the foundational component for building high impact customer features. Furthermore, piloting proves value and de risks innovation before a full scale deployment.

By applying targeted AI solutions on a disciplined GCP architecture, we are successfully scaling operations, reducing manual effort, and enhancing the customer journey through smarter presentation and discovery tools. This is just the beginning; we are continuously pushing for a more intelligent, fully optimized e-commerce platform.

Jonny Johansson, Data System Director & Steffan Mølbæk Andersen, Data Intelligence Director