MonitorIQ

Draft-only AI appeal review for Reddit moderators: risk flags, proof gaps, safe reply drafts, and team handoffs without automatic moderation actions.

Comment

Inspiration

Ban appeals are one of the hardest moderation workflows to scale. A moderator often has to read a messy modmail thread, reconstruct what happened, compare it against subreddit rules, check whether proof is missing, decide if the appeal is safe to continue, and then write a careful reply that does not leak internal notes.

MonitorIQ was built to make that workflow feel more like an intelligence dashboard for moderators: a place where a mod team can scroll through appeals, see what matters first, ask an agent for help, and move from confusion to a clear next action without giving automation control over Reddit decisions.

What it does

MonitorIQ turns ban appeals into structured review packets for moderators. Instead of forcing a team to read every appeal from scratch, it organizes each case into:

  • a concise appeal summary;
  • risk flags for threats, harassment, doxxing, spam, legal threats, self-harm, and ban-evasion admissions;
  • missing proof or clarification needed from the user;
  • likely subreddit rule matches;
  • safe draft modmail language;
  • model/source status with deterministic fallback notes;
  • moderator outcome controls that record notes without taking automatic Reddit actions.

The local demo also includes an agentic moderator assistant. A moderator can ask questions like “Which appeal should I open first?”, “Prepare today’s handoff,” or “How should we handle this inflammatory appeal?” The assistant responds with practical triage, proof gaps, safety routing, and draft language grounded in the visible case queue.

The goal is the same feeling as a good creator analytics tool or operations dashboard: scroll, inspect, understand the priority, and act faster. But MonitorIQ is designed specifically for moderation judgment, not content growth.

Moderator-first design

MonitorIQ is intentionally human-in-the-loop.

It does not automatically unban users, uphold bans, remove posts, archive modmail, mute users, or send replies. Draft replies stay drafts. Private moderator notes are not copied into user-facing language. Safety-sensitive appeals are routed toward closure or second review instead of normal appeal handling.

The app’s job is to reduce moderator load, not replace moderator authority.

How we built it

MonitorIQ is built with Reddit Devvit, TypeScript, Redis-backed app storage, Devvit menu actions, a custom dashboard, and a local demo environment.

The analysis system supports optional server-side OpenAI calls, while keeping the API key off the browser. If a model route is unavailable or returns invalid JSON, the app falls back to deterministic local triage so the moderation workflow still works during demos and review.

We also built dataset and evaluation assets around the appeal-review schema: final train, validation, locked gold eval, and norm-holdout records. Stress artifacts cover multi-turn agent chat, queue visualization, case opening, daily reports, export behavior, mobile layout, and clean console checks.

Challenges

The main challenge was making the assistant useful without making it dangerous. Ban appeals are high-context and sometimes emotionally charged, so the product needed clear safety boundaries: no automatic enforcement, no hidden final decisions, no leaking private notes, and no pretending uncertain cases are certain.

We also ran into Devvit playtest friction around newly created test subreddits. To keep the project reviewable, we focused on a reliable uploaded app listing, a runnable local demo, seeded review cases, dataset validation, and stress-tested workflows that demonstrate the intended moderator experience end to end.

What we learned

The best moderation tools do not just classify content. They reduce context switching. They help moderators see what changed, what is risky, what proof is missing, and what reply would be safe.

MonitorIQ turns appeal handling from a scattered modmail reading task into a focused review workflow: triage, inspect, draft, hand off, and decide with the moderator still in control.

Built With

Share this project:

Updates