Understanding the challenge

Over a million messages every day, across multiple brands, operators, and sports.

The challenge wasn't simply handling volume. Moderators needed to quickly understand the context of a message, assess its severity, review previous user behaviour and decide what action to take, without losing track of where the content originated.

At the same time, the system needed to work across multiple brands, operators and sports while supporting three user roles with different levels of access and control.

My challenge was to turn that operational complexity into a moderation experience where the right information and actions were available at the right moment — while creating a system that could scale as Virtual Stadium grew.

Designing moderation for a unique industry

I started by analyzing established moderation tools and patterns from other industries to understand how high-volume platforms surface reports, user history and moderation actions.

While moderation tools already existed, none were designed specifically around the dynamics of social betting at this scale. It introduced risks those systems weren't designed around. Harmful behaviour wasn't limited to offensive language or spam, moderators also needed to consider betting-specific behaviours such as misleading tips and attempts to manipulate odds.

For the MVP, we prioritized the highest-value moderation workflows while creating an architecture that could expand as new client requirements, behaviors and moderation rules emerged.

Bringing AI into the moderation workflow

With millions of messages passing through the platform each day, relying entirely on manual moderation wasn't scalable. But fully automating moderation introduced a different risk: context matters, and not every decision can be reduced to a score.

Automating repetitive moderation without removing human control

Starting from zero gave us the opportunity to define where AI could genuinely improve the moderation process. We developed two levels of AI support based on how much context and human judgement a decision required.

• Autopilot evaluates messages against predefined moderation rules, taking into account offensiveness scores, previous user infractions, and user reports.

• Copilot supports moderators with AI-generated offensiveness scores, enabling semi-automated moderation of clearly offensive or inoffensive content.

The wider moderation system also supports configurable AI thresholds, automated content filtering, blacklisted words, spam detection, and automatic sanctioning rules.

The focus was on using AI to support moderation at scale while keeping moderators in control of key decisions.

Designing for clarity and control

The interface needed to make complex information easier to process and important actions easy to access.

One of the key challenges was designing a hybrid navigation system that could support a growing set of features. Message display also required careful consideration, as content could come from multiple brands, operators, and sports.

The platform needed to support three user levels with different responsibilities and access requirements. Giving every user access to the entire system would have introduced unnecessary complexity and increased the risk of actions being taken by the wrong people.

We defined permissions around each role's responsibilities, and I translated them into the product experience by controlling which areas, information and actions were available at each level.

The solution

The result is a scalable, AI-supported moderation system designed specifically for social betting environments.

By combining automated moderation with clear human oversight, the platform helps moderation teams process high volumes of messages, review flagged content, manage users, and maintain safer online communities across multiple brands and operators.

The system balances automation, complexity, and human control, turning over a million daily messages into structured, manageable moderation workflows.

My takeaways

Getting comfortable designing without all the answers.

Starting from zero meant that many parts of the product were still undefined when the design process began. I learned to work comfortably with ambiguity — turning assumptions into testable concepts, making informed decisions with what we knew at the time, and adapting the system as we learned more about our users, technical constraints and client needs.