How we integrated a Vision Model to detect illegal postings in an E-commerce Market Place

How AI implemented correctly can help staff in day to day processes.

Client Overview

Industry: E-commerce

Company Size: 20+ Employees

Main Offer/Product: B2C Marketplace providers

Users: 2+ million; 2,000+ subscribed vendors

Initial Situation & Challenges

Our client has been operating in the e-commerce industry since 2014, successfully managing and scaling their marketplace platform with the support of a dedicated in-house team. The platform serves two categories of vendors, subscribed (verified) and unsubscribed (unverified). Both of these require distinct moderation workflows.

To maintain compliance and safeguard platform integrity, every post submitted by unsubscribed vendors had to undergo manual verification before publication. In parallel, the moderation team continuously reviewed content from subscribed vendors to ensure ongoing adherence to the platform’s Terms of Service (TOS) and to prevent the circulation of prohibited or illegal material.

At the time of our collaboration, the client maintained a team of over seven moderators responsible for manually screening and approving thousands of posts daily, a process that had become increasingly resource-intensive as the platform scaled.

Key Pain Point:

  1. The extremely high workload needed to maintain the marketplace with only the posts that adhere to the TOS, while also making sure to process the posts timely to not risk losing vendors. This emerged as a major pain point as the platform experienced significant growth in the two years leading up to our collaboration.

Our Approach & Actions Taken

Analysis & Insights:

To understand the processes of moderation we requested to be present during administration hours, where we got to study the end-to-end operations in detail and receive direct training on how both subscribed and unsubscribed vendors were managed. We were also provided the TOS and everything that our integration aimed to keep off the marketplace, to mention a few: images of children, animals, documents with private information (such as social security numbers), weapons, etc.

Requirements & Expectations

The goal was to centralize and streamline the moderation process by creating a unified dashboard where all potentially illegal posts could be automatically highlighted and displayed in one place. This allowed the moderation team to efficiently review each flagged post and take the appropriate action, either delete the post if it violated the Terms of Service or approve it if it was deemed compliant.

In parallel, the system was designed to collect feedback from moderators on each AI-generated highlight, indicating whether it was a correct or incorrect detection. This feedback loop provided essential data to continuously improve the accuracy and reliability of the Vision Agent’s detection model.

As the system matured, the next phase involved extending automation capabilities, not only highlighting suspicious posts but also identifying repeat-offender accounts for potential suspension and automatically removing posts or images that were confidently detected as violating the platform’s policies.

Integration with Gemini Vision Model

To bring this system to life, we integrated Google’s Gemini Vision Model, capable of understanding both text and image content. This allowed our moderation framework to analyze posts in depth and not just based on captions or titles, but also on the actual visual elements within the images uploaded by vendors.

We designed the workflow so that every new post submitted by an unsubscribed seller was automatically processed through our pipeline. The text content (title, description) and all accompanying images were sent to the Gemini Vision API for evaluation. The model then returned a structured response that classified the content according to risk categories defined in collaboration with the client, “Approved” or “Rejected”.

To ensure scalability and precision, we optimized the detection logic using parallel batch image analysis, allowing multiple images to be processed in a single request. For each flagged item, the system generated confidence scores and highlighted the specific reasons behind the detection (e.g., presence of weapons, or personal documents).

The results were then centralized into the flagged-post dashboard, where staff could review AI findings, approve or delete posts, and provide feedback. Over time, this feedback loop was used to refine the model’s prompts and improve contextual understanding. The combination of Gemini’s visual intelligence and our feedback-driven iteration process resulted in a highly reliable, self-improving moderation system capable of maintaining compliance at scale with minimal manual effort.

Closing Results & Impact

The integration of our Vision Model System transformed the client’s moderation workflow from a fully manual process into a semi-automated, AI-assisted system that drastically reduced the team’s operational burden and improved the overall consistency of content moderation. What once required a team of seven moderators working around the clock became a streamlined process where AI handled the initial screening, allowing human moderators to focus on more strategic and high-value tasks.

This project demonstrated how AI-driven vision systems, when implemented with structured feedback loops and human oversight, can help organizations achieve the perfect balance between efficiency and accuracy, ultimately empowering teams to work smarter and focus on what truly drives long-term growth.

Share:

LinkedIn

More Posts