DesiMode AI

AI-powered multilingual moderation assistant for Reddit communities. It detects toxicity, spam, and policy violations and helps moderators review and manage posts faster and more efficiently.

Comment

DesiMode AI — Multilingual AI Moderation for Reddit Communities

Inspiration

Reddit communities in South Asia often face moderation challenges that traditional moderation tools are not designed to handle effectively. Toxicity, scams, spam, and harassment are frequently written in mixed languages, transliterated text, slang, or regional dialects that bypass standard keyword-based moderation systems.

For example, harmful content may appear as:

  • Hindi written using English characters (Hinglish)
  • Bangla written using Romanized text (Banglish)
  • Mixed English + regional language abuse
  • Regional scam phrases and slang variations

Many moderation bots are optimized mainly for English-speaking communities, leaving multilingual communities underserved.

DesiMode AI was inspired by the need to create a smarter moderation system that understands multilingual community behavior while still keeping moderators fully in control of escalation decisions.

My goal was to build a moderation tool that could reduce moderator workload, improve moderation quality, and support communities that are often overlooked by existing AI moderation systems.

What it does

DesiMode AI is an AI-powered moderation system built on Reddit Devvit.

The tool automatically analyzes Reddit posts and comments to detect:

  • Toxicity
  • Harassment
  • Spam
  • Scam attempts
  • Harmful multilingual content

The app supports:

  • English
  • Bangla
  • Hindi
  • Hinglish
  • Banglish
  • Mixed-language content

Core Features

  • AI-powered moderation using Gemini 3.1 Flash Lite
  • Multilingual toxicity and scam detection
  • Context-aware moderation instead of keyword-only filtering
  • Automated 3-strike enforcement workflow
  • Warning messages and ModNote logging
  • Structured ModMail escalation system
  • Moderator dashboard with near real-time analytics
  • Moderator-only dashboard access
  • Real-time dashboard updates using Devvit Realtime

The moderation system follows a balanced moderation philosophy:

  • harmful content is handled automatically
  • moderators remain in control of final escalation decisions
  • no automatic banning is performed

How I built it

DesiMode AI was built using Reddit's Devvit platform with a custom Devvit Web dashboard architecture.

Tech Stack

  • TypeScript
  • Reddit Devvit
  • Devvit Web
  • Redis
  • Google Gemini 3.1 Flash Lite
  • Devvit Realtime APIs
  • Node.js

Moderation Workflow

The moderation pipeline works like this:

  1. A Reddit post or comment is submitted
  2. The content is analyzed using Gemini AI
  3. The AI classifies the content:
  • Safe
  • Toxic

  • Scam / Spam

    1. Strike history is checked in Redis
    2. Enforcement actions are triggered automatically:

  • content removal

  • warning messages

  • ModNote logging

  • moderator escalation after 3 strikes

I also built a custom moderator dashboard using Devvit Web to provide:

  • moderation statistics
  • escalation tracking
  • violation analytics
  • estimated moderator time saved
  • near real-time updates

To support reliable dashboard refreshes without using socket.io, we combined lightweight polling with Devvit Realtime events.

Challenges I ran into

Multilingual moderation complexity

One of the biggest challenges was handling multilingual and transliterated content accurately.

Traditional keyword matching struggled with:

  • Hinglish
  • Banglish
  • slang
  • transliteration
  • intentional spelling manipulation

I was needed a more contextual moderation system instead of relying purely on static keyword lists.

Devvit architecture migration

Reddit shifted away from older Blocks-based approaches toward Devvit Web.

I have redesigned parts of the app architecture to use:

  • custom webviews
  • Devvit Web components
  • realtime synchronization

This required restructuring parts of the frontend and dashboard system.

AI reliability and moderation safety

AI moderation systems can easily become overly aggressive or inaccurate.

I wanted to ensure:

  • moderators stay in control
  • escalation is transparent
  • the system remains Reddit policy compliant

Because of this:

  • DesiMode AI does not auto-ban users
  • moderators receive structured escalation summaries
  • AI reasoning and confidence scores are included

Real-time dashboard synchronization

Building a near real-time moderation dashboard without traditional websocket frameworks required careful optimization using Devvit Realtime and efficient polling strategies.

Accomplishments that I am proud of

I am are especially proud of:

  • Building multilingual moderation support for underserved Reddit communities
  • Successfully supporting English, Bangla, Hindi, Hinglish, and Banglish moderation workflows
  • Creating a production-oriented Devvit moderation system
  • Designing a structured 3-strike escalation workflow
  • Building a moderator-only secure dashboard
  • Implementing AI reasoning-based moderation explanations
  • Creating a scalable moderation history system using Redis
  • Maintaining moderator control instead of relying on unsafe full automation

I am also proud that the project focuses on real moderation problems faced by actual multilingual communities instead of only building a prototype demo.

What I learned

This project taught me several important lessons:

  • Real moderation systems require contextual understanding, not just keyword filtering
  • Multilingual communities face unique moderation challenges that are often underserved
  • Reliability and transparency are more important than aggressive automation
  • Moderator trust is critical for moderation tools
  • Devvit provides powerful infrastructure for building native Reddit moderation experiences
  • AI moderation systems work best when humans remain part of the escalation process

I am also learned a great deal about:

  • realtime system design
  • moderation workflows
  • Devvit architecture
  • scalable backend state management
  • AI safety considerations

What's next for DesiMode AI

I have plan to continue improving DesiMode AI with:

  • Additional language support
  • Better scam pattern detection
  • Community-specific moderation tuning
  • Moderator customization settings
  • Improved moderation analytics
  • Smarter false-positive reduction
  • Better onboarding and installation flows
  • Cross-community moderation insights

My long-term goal is to make multilingual moderation significantly easier for Reddit communities across South Asia and other multilingual regions.

I hope DesiMode AI can help moderators spend less time fighting spam and abuse — and more time building healthy communities.

Built With

Share this project:

Updates