AuraMod — AI-Powered Proactive Reddit Moderation

Inspiration

• Mods always find flame wars after the damage is done
• Existing tools are reactive — we wanted proactive detection
• Key insight: toxicity is about patterns, not just single comments

How It Works

• Velocity Tracking — Redis sorted sets count replies per thread in a 10-min sliding window
• AI Analysis — When 3+ replies hit in 10 min, OpenRouter AI (Gemini 2.0 Flash) scores toxicity from 0.0 (toxic) → 1.0 (safe)
• Auto-Intervention — Lock, remove, or report based on configurable thresholds
• Cooldown Guard — 30-min Redis TTL prevents duplicate actions on same thread
• Keyword Fallback — Works without an API key using tiered keyword heuristics

Mod Tools

• 🔍 Analyze Comment — Right-click menu for instant AI toxicity score
• 📋 Intervention Log — View recent automated actions per subreddit
• 📧 Modmail Alerts — Auto-notifies mod team with score + context

Challenges Faced

• Nested replies broke velocity tracking → fixed by tracking per-post instead of per-parent
• Bot re-triggered repeatedly on hot threads → added Redis cooldown system
• AI sometimes returns text instead of numbers → built regex extraction + multi-layer fallback

What We Learned

• Redis sorted sets are ideal for time-windowed counting
• LLM outputs need guardrails and fallback chains
• Devvit triggers are stateless — Redis becomes your app's memory

Built With

• Platform: Reddit Devvit (TypeScript)
• AI: OpenRouter API → Google Gemini 2.0 Flash
• Storage: Devvit Redis (velocity, cooldowns, intervention logs)

Built With

• devvit
• node.js
• reddit
• typescript