AuraCouncil

• 

  the analyzer

• 

  input

Inspiration

Moderating a large subreddit is emotionally draining. Much of the workload comes from users unintentionally breaking complex rules or posting low-effort content. We wanted to build a system that shifted moderation from a punitive action (removing a post and banning the user) to an educational one (coaching users to improve their posts before they submit). We were inspired by the idea of a "co-pilot" for healthy community engagement.

What it does

AuraCouncil acts as a pre-emptive "Submission Pilot." Instead of the standard Reddit text box, users draft their posts in our Devvit WebView app. A multi-agent "Mod Council" (comprising the Compliance Officer, Tone Guardian, Engagement Strategist, and Media Inspector) instantly evaluates the draft across four dimensions: rule compliance, tone, engagement quality, and media safety. It scores the draft and provides smart, actionable suggestions. A post cannot be submitted until it reaches the moderator-defined safety threshold.

How we built it

We utilized Reddit's Devvit platform for both the frontend and backend. The UI is built with React in a Devvit Custom Post WebView, styled with Tailwind CSS for a premium, glassmorphism aesthetic. The backend relies on @devvit/web/server and tRPC to handle complex state securely.

The core intelligence is driven by Google's Gemini 3 family (Gemini 3 Flash Preview & Gemini 3.1 Flash). We engineered a complex JSON schema prompt that forces Gemini to act as a council of distinct agents, each providing granular reasoning and status codes. For advanced accuracy, the app dynamically queries Devvit's Reddit APIs to fetch recent top posts (as positive examples) and recently removed posts (as negative examples), injecting a dynamic few-shot learning context into the model prompts. We also utilized Devvit's native Redis plugin to build a powerful Moderator Dashboard, allowing mods to inject custom subreddit rules and thresholds dynamically, and to track live time-savings analytics.

Challenges we ran into

• Devvit Fetch Timeouts: Initially, allowing users to upload 20 images caused DEADLINE_EXCEEDED errors because the internal Devvit fetch queue couldn't process the Gemini API response within the platform's strict execution limits. We solved this by capping the image limit to 3 and utilizing client-side canvas compression, ensuring absolute stability.
• Strict Typing for AI Responses: Parsing raw LLM output into a dynamic React UI was brittle. We solved this by using Gemini's responseSchema parameter to enforce strict JSON schemas, guaranteeing our UI always received structured arrays for the agent logs and auto-fix suggestions.

Accomplishments that we're proud of

• Building a genuinely useful mod tool that addresses the root cause of mod fatigue rather than just band-aiding the symptoms.
• Designing a beautiful, non-intimidating user interface that feels native to Reddit but leverages state-of-the-art AI.
• Successfully integrating Devvit's Redis to make the app dynamically configurable per-subreddit, fulfilling the true definition of a scalable "Mod Tool".

What we learned

We learned the intricacies of building within the Devvit ecosystem, specifically handling WebView limitations and optimizing API calls to respect Devvit's execution time limits. We also discovered the immense power of system instructions and schema enforcement when working with the Gemini 3 family to guarantee reliable, structured output.

What's next for AuraCouncil

In the future, we want to integrate real-time Reddit API data into the Mod Council's context—such as the user's past ban history or current subreddit trending topics—to provide even more personalized coaching. We also plan to allow moderators to customize the personalities of the AI agents to better fit the vibe of their specific communities.

Built With

• gemini
• python