AMIS — AI Moderation Intelligence System

AMIS is an AI-powered moderation platform that combines embeddings + LLMs to automate Reddit moderation in real time while reducing AI costs by 60–70% through intelligent routing.

Comment

AMIS — AI Moderation Intelligence System

Inspiration

Online communities are growing faster than human moderation teams can scale. Most moderation tools today are either simplistic keyword filters that fail on nuance or expensive AI systems that send every request through an LLM, making them costly and difficult to scale.

We wanted to explore whether it was possible to build a real-time AI moderation system that was:

  • fast,
  • cost-efficient,
  • explainable,
  • and deployable entirely inside Reddit’s constrained Devvit serverless environment.

The challenge became even more interesting because Devvit imposes strict limitations:

  • 3-second execution windows,
  • KV-only persistence,
  • no traditional backend servers,
  • limited UI capabilities,
  • strict rate limits.

Instead of treating those as blockers, we designed the entire architecture around those constraints.

What it does

AMIS (AI Moderation Intelligence System) is an AI-powered Reddit moderation platform that helps moderators analyze and triage content in real time.

The platform:

  • ingests subreddit posts automatically,
  • compares them against subreddit policies,
  • performs semantic analysis using embeddings + LLMs,
  • and surfaces moderation recommendations through a custom live dashboard embedded directly inside Reddit.

Key capabilities include:

  • real-time moderation analysis,
  • embedding-based semantic policy matching,
  • AI-assisted moderation recommendations,
  • duplicate content detection,
  • moderator consistency analysis,
  • persistent moderation history,
  • multi-provider AI support,
  • cost-aware LLM routing,
  • background queue orchestration.

AMIS supports:

  • OpenAI,
  • Google Gemini,
  • Anthropic Claude,
  • and local/self-hosted OpenAI-compatible models like Ollama.

How we built it

The system was built entirely on Reddit’s Devvit platform using TypeScript and a serverless architecture.

Core Architecture

We designed a two-tier AI moderation pipeline:

Embedding Similarity → Cheap Decision
Uncertain Cases → LLM Escalation

Every submitted post goes through:

  1. embedding generation,
  2. cosine similarity matching against subreddit rules,
  3. confidence evaluation,
  4. optional LLM escalation,
  5. recommendation generation.

Cost-Aware AI Routing

One of the biggest engineering goals was minimizing unnecessary LLM calls.

We implemented threshold-based routing:

similarity < 0.45  → auto approve
similarity > 0.85  → auto flag
0.45–0.85          → LLM classification

This reduced LLM API usage by roughly 60–70%.

Background Processing

Since Devvit UI actions timeout after 3 seconds, we moved long-running tasks into scheduled background jobs using Devvit Scheduler.

We built:

  • queue processors,
  • retry systems,
  • exponential backoff handling,
  • async-safe workflows,
  • cache layers,
  • stale-data filtering.

Dashboard & UX

We also built a fully interactive moderation dashboard directly inside Reddit using Devvit Block Kit.

Features include:

  • live moderation queue,
  • auto-refreshing views,
  • moderation history,
  • consistency insights,
  • live countdown banners,
  • filter chips,
  • pagination.

Challenges we ran into

Platform Constraints

The biggest challenge was engineering inside a heavily constrained serverless environment:

  • no traditional backend,
  • no database,
  • KV-only storage,
  • hard execution limits,
  • fixed-height UI constraints,
  • no native scrolling.

Many common architectural approaches simply were not possible.

We had to redesign:

  • queue orchestration,
  • state synchronization,
  • async execution,
  • caching,
  • and UI rendering specifically for Devvit.

AI Reliability

Moderation is difficult because rules vary dramatically.

Some rules are structural:

  • “Titles cannot start with A”

Others are semantic:

  • “No misinformation”

We had to engineer prompts and fallback systems capable of handling both categories reliably.

We also built:

  • resilient JSON parsing,
  • markdown cleanup,
  • fallback classification,
  • embedding-only recovery flows.

Cost Optimization

Naively calling an LLM for every post would have made the platform extremely expensive.

Designing an intelligent routing system that balanced:

  • cost,
  • accuracy,
  • speed,
  • and moderation quality

became one of the hardest and most rewarding engineering problems in the project.

Accomplishments that we're proud of

  • Built and deployed a production AI moderation platform entirely inside Reddit’s Devvit ecosystem
  • Reduced LLM usage by ~60–70% using embedding-based routing
  • Achieved near real-time moderation analysis (<5 seconds)
  • Successfully orchestrated distributed AI workflows without dedicated backend servers
  • Built a multi-provider AI abstraction layer supporting OpenAI, Gemini, Claude, and local models
  • Designed a moderation dashboard with live updates and persistent history
  • Implemented moderator consistency analysis and duplicate detection systems
  • Engineered reliable background queue processing inside strict serverless constraints

What we learned

This project taught us a huge amount about:

  • distributed systems,
  • AI infrastructure,
  • vector embeddings,
  • prompt engineering,
  • serverless orchestration,
  • async workflows,
  • reliability engineering,
  • and platform-aware architecture.

More importantly, we learned that production AI systems are far more about:

  • orchestration,
  • caching,
  • retries,
  • latency,
  • reliability,
  • and cost optimization

than simply calling an LLM API.

We also learned how much thoughtful system design matters when operating under hard platform constraints.

What's next for AMIS — AI Moderation Intelligence System

We want to evolve AMIS from a Reddit moderation assistant into a generalized AI moderation and policy intelligence platform.

Future plans include:

  • cross-platform moderation support,
  • smarter agentic moderation workflows,
  • adaptive moderation learning,
  • moderator behavior analytics,
  • improved explainability,
  • real-time policy evolution,
  • collaborative moderation intelligence,
  • and scalable trust & safety infrastructure.

Long term, we envision AMIS becoming an AI-native moderation layer for communities, marketplaces, forums, and collaborative platforms.

Built With

  • async
  • claude
  • cosine-similarity-search
  • devvit-block-kit
  • devvit-kv-store
  • devvit-scheduler
  • gemini
  • jest
  • ollama
  • openai-gpt-4o
  • queue
  • reddit-devvit
  • serverless-architecture
  • typescript
  • vector-embeddings
Share this project:

Updates