Inspiration I’ve spent a lot of time in Discord communities and noticed a recurring problem: traditional moderation bots are easily tricked by simple spelling bypasses. Moderators end up in a constant game of "whack-a-mole" with toxic users. I wanted to build something that doesn't just look for bad words but actually understands the intent of a conversation.
What it does Murdoch is a high-efficiency semantic moderation engine. It uses a three-layer pipeline to catch everything from blatant slurs to subtle harassment and social engineering. By combining fast local filtering with AI-driven context analysis, it keeps servers safe without the lag or high costs of typical AI bots.
How we built it The core engine is built with Rust for maximum performance and memory safety.
Layer 1: A lightning-fast Regex filter for instant blocks.
Layer 2: A custom message buffer to batch data and reduce API latency.
Layer 3: Gemini 3 Flash for deep semantic analysis of chat context. The frontend dashboard is built to handle real-time logs and user appeals.
Challenges we ran into
The biggest hurdle was managing the asynchronous flow between the Discord gateway and the Gemini API. Implementing a "double-buffering" system in Rust was tricky—I had to dive deep into Arc, Mutex, and async channels to ensure the bot could process thousands of messages without blocking or losing data.
Accomplishments that we're proud of
I’m really proud of the stability. With over 250 tests backing it up, the bot is rock-solid. Even if the AI API goes down, the Regex layer keeps working, so the server is never left unprotected. It’s production-grade software ready to scale.
What we learned
I learned that Rust is incredibly fast and its compiler is a great mentor—those error messages make debugging much easier than in other languages. I also gained a much deeper understanding of how to use AI as a "deterministic-first" tool rather than just a black box.
What's next for Murdoch
I plan to add more granular server-specific "personalities" so admins can tune the bot's strictness. I’m also looking into adding image-based moderation using Gemini's multimodal capabilities to catch toxic memes and phishing screenshots.
Log in or sign up for Devpost to join the conversation.