Ilan Zarantonello
Visual Production & Innovation
AI-Powered Image Quality Moderation
This project was initiated to solve a critical and growing challenge within a major e-commerce content pipeline: ensuring consistent image quality at scale.
My Role
As the Project Lead, I identified the inefficiencies of the manual quality check process, framed the problem with data, and am currently leading the discovery and vendor evaluation process to implement an AI-powered solution.
The Challenge: Inconsistency and Inefficiency at Scale
After our team successfully improved our lighting, style guides, and retouching guidelines, we needed a scalable way to protect those quality gains. The existing manual quality check process was a major bottleneck.
- Time-Consuming: Manual checks required significant effort from multiple teams, including Quality Management, Art Directors, and Retouching , consuming approximately 3.5 hours per day.
- Limited Scope: Due to the high volume of content, across different vendors and partners, we could only manually review an estimated 6% of the nearly 1,000 items produced daily , leaving 94% of SKUs unchecked.
- Inconsistent Results: Reliance on manual checks led to inconsistencies in image quality across different vendors and internal teams , which can negatively impact customer experience and brand image.
The Discovery Process: A Data-Driven Approach
To build a strong business case, I initiated a thorough discovery phase. The goal was to move from anecdotal evidence to quantifiable data.
- Stakeholder Interviews: I conducted talks with key teams—including Quality Management, Retouching, and Art Directors —to understand their current workflows, identify their primary pain points , and gather their ideas.
- Quantifying the Impact: I gathered data to measure the time invested in manual checks , the number of errors being reported, and the ratio of checked vs. unchecked assets. This provided a clear baseline to measure the potential ROI of a new solution.
The Proposed Solution: Centralised AI Moderation
My proposed solution is to implement a centralised, AI-powered tool to automate quality checks for all packshots. This will ensure consistent standards, reduce manual effort, and allow teams to focus on more strategic work.
We idientified few possible vendors that could help us customising an AI tool for us.
Key Features and Requirements
- Detect adherence to required file sizes, formats, and resolutions.
- Identify consistent and clean backgrounds (color, uniformity).
- Perform blur detection.
- Detect dirt, unwanted marks, dust, or imperfections.
- Ensure product symmetry where applicable.
- Verify correct proportions of the product relative to the background.
- Reliably detect defined error types, including: colour/light inconsistencies, wrong colour (image vs description), different croppings (subject position size), cropped subject, uploads errors (bastelbild, wrong order), packshots styling (flatlays vs mannequin), wrong file requirements, low image quality (blur, pixelated), and retouch mistakes.
- Ability to integrate with existing systems, such as upload platforms and databases.
- Functionality for screening images before and/or after they are online.
- Customisable templates and rules for specific e-commerce image guidelines.
- Capable of handling high volumes of images and increasing business needs.
- Build a foundation for future expansion to model imagery, pending legal frameworks.
- Potential for detecting model pose, garment fit, and lighting.
Measuring Success: The Pilot Scorecard
The success of the pilot will be measured against clear, predefined metrics:
- Error Detection Accuracy: We will measure the tool's Precision and Recall for key error types to ensure its reliability.
- Workflow Efficiency: We aim to see a significant reduction in the time spent on manual quality checks per SKU.
- Cost-Effectiveness: We will calculate the potential ROI based on time saved and the cost of implementation.
Current Status:
This project is currently in the final stages of the Discovery & Vendor Decision phase.