# Glin-Profanity Roadmap
This document outlines the future direction and planned features for glin-profanity.
## Vision
Glin-profanity aims to be the most comprehensive, developer-friendly, and AI-native content moderation library available. We're building for a future where AI assistants and applications need reliable, fast, and intelligent profanity detection.
---
## Market Opportunity
Content moderation is a **$14B market in 2026**, growing to **$42B by 2035** (13% CAGR).
- Cloud/SaaS deployment: 70% of market by 2035
- Key competitors: OpenAI Moderation (free), Azure Content Safety, Sightengine, Hive
### Our Differentiators
| Feature | OpenAI | Azure | Sightengine | glin-profanity |
|---------|--------|-------|-------------|----------------|
| Languages | 1 | Few | Few | **24** |
| Self-hosted | ❌ | ❌ | ❌ | **✅** |
| Open Source | ❌ | ❌ | ❌ | **✅** |
| Leetspeak | ❌ | ❌ | ✅ | **✅** |
| AI Framework SDKs | N/A | N/A | ❌ | **✅** |
| Price | Free | $0.38/1K | Paid | **Free** |
---
## Technical Architecture
```
┌─────────────────────────────────────────────────────────────────┐
│ HYBRID DETECTION ARCHITECTURE │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Input Text │
│ ↓ │
│ ┌─────────────────────────────────────────────┐ │
│ │ Layer 1: Dictionary (Core - Always Fast) │ ~1ms │
│ │ • 24 language word lists │ │
│ │ • Leetspeak normalization │ │
│ │ • Unicode homoglyph detection │ │
│ │ • Context-aware filtering │ │
│ └─────────────────────────────────────────────┘ │
│ ↓ (optional, for edge cases) │
│ ┌─────────────────────────────────────────────┐ │
│ │ Layer 2: ML Model (Optional - Accurate) │ ~50-100ms │
│ │ • transformers.js + ONNX │ │
│ │ • Pre-trained: pardonmyai, toxic-bert │ │
│ │ • Context understanding │ │
│ └─────────────────────────────────────────────┘ │
│ ↓ │
│ Combined Result (confidence score + detected words) │
│ │
└─────────────────────────────────────────────────────────────────┘
```
---
## Current State (v3.2.0 + MCP v1.2.0 + AI Integrations)
- 24 language support
- Leetspeak detection (basic, moderate, aggressive)
- Unicode homoglyph normalization
- Context-aware filtering
- React hooks support
- Optional ML/TensorFlow integration
- **MCP Server v1.2.0** with:
- 19 tools for content moderation
- 5 workflow prompts
- 5 reference resources
- Conversation memory (user tracking)
- Real-time streaming support
- HTTP transport for cloud deployment
- **AI Framework Integrations** (NEW):
- OpenAI function calling tools (`glin-profanity/ai/openai`)
- LangChain tools (`glin-profanity/ai/langchain`)
- Vercel AI SDK tools + middleware (`glin-profanity/ai/vercel`)
- Semantic analysis with embeddings (`glin-profanity/ai/semantic`)
---
## Q1 2026: AI-Native Features
### MCP Server Enhancements
- [x] **Streaming support** - Real-time profanity detection for chat streams ✅
- [x] **Conversation memory** - Track user behavior patterns across messages ✅
- [x] **Smart prompts** - Pre-built prompts for common moderation workflows ✅
- [x] **HTTP transport** - Deploy as a remote MCP server for cloud scenarios ✅
### AI Integration APIs
- [x] **OpenAI function calling** - Direct integration with GPT models ✅
- [x] **LangChain tool** - Native LangChain integration ✅
- [x] **Vercel AI SDK** - Middleware for AI applications ✅
- [x] **Semantic analysis hooks** - Combine with embeddings for semantic moderation ✅
---
## Q2 2026: Advanced Detection & ML
### Multi-Modal Support
> **Strategy**: BYO (Bring Your Own) approach - keep core lightweight, optional integrations
- [ ] **Image text extraction (OCR)** - `glin-profanity/ocr` subpath with Tesseract.js
- Approach: Optional peer dependency, users install if needed
- Bundle impact: 0 in core, ~5MB if OCR module imported
- [ ] **Audio transcription integration** - `glin-profanity/audio` subpath
- Approach: Accept transcribed text from Whisper/Google STT/etc.
- NOT bundling Whisper - just provide pipeline utilities
- [ ] **Meme detection** - DEFERRED (complex, low ROI for now)
### ML-Enhanced Detection (Hybrid Approach)
> **Strategy**: Dictionary-first (fast), ML-second (accurate when needed)
> Based on research: [Detoxify](https://github.com/unitaryai/detoxify), [transformers.js](https://huggingface.co/docs/transformers.js)
- [ ] **transformers.js integration** - Optional ONNX model support
- Use existing models: [pardonmyai](https://huggingface.co/tarekziade/pardonmyai) (97.5% accuracy)
- Fallback to dictionary when ML unavailable
- [ ] **Confidence scoring** - ML provides probability, dictionary provides certainty
- [ ] **Custom model training guide** - Documentation for fine-tuning
- [ ] **Hugging Face Space** - Interactive demo for market visibility
### Language Improvements
> **Note**: Dictionary approach works well for 24 languages. ML models (XLM-RoBERTa) could improve
> low-resource languages by +23% but require significant model size (~500MB).
- [ ] **Improve word lists** - Community contributions for existing 24 languages
- [ ] **Add high-demand languages** - Vietnamese, Indonesian, Tagalog (based on npm usage)
- [ ] **Regional variants** - British vs. American English profanity
- [ ] **Slang database** - Gen-Z, internet culture, gaming terms
### Enhanced NLP
- [ ] **Sentiment + profanity combo** - Is it angry profanity or friendly banter?
- [ ] **Intent classification** - Threat detection vs. casual swearing
- [ ] **Toxicity gradients** - Severity levels with ML confidence
---
## Q3 2026: Enterprise Features
### Moderation Workflows
- [ ] **Queue management** - Moderation queue API
- [ ] **Appeal handling** - False positive reporting and learning
- [ ] **Audit logging** - Detailed logs for compliance
- [ ] **Rule builder** - Visual rule creation interface
### Performance & Scale
- [ ] **Edge deployment** - Cloudflare Workers, Vercel Edge
- [ ] **WebAssembly** - WASM build for browser-native speed
- [ ] **Redis caching** - Distributed caching for high throughput
- [ ] **Batch API** - 10,000+ texts per second processing
### Compliance
- [ ] **GDPR tools** - Data handling compliance helpers
- [ ] **Content policies** - Pre-built policy templates (COPPA, CIPA, etc.)
- [ ] **Reporting dashboards** - Analytics and metrics
- [ ] **Export formats** - Compliance report generation
---
## Q4 2026: Ecosystem Expansion
### Framework Integrations
- [ ] **Next.js middleware** - Server/edge middleware component
- [ ] **Remix loader** - Native Remix integration
- [ ] **SvelteKit hooks** - Svelte integration
- [ ] **Vue composables** - Vue 3 composition API
- [ ] **Astro integration** - Astro middleware
### Platform SDKs
- [ ] **Discord.js plugin** - Bot integration
- [ ] **Slack app** - Workspace moderation
- [ ] **Twitch extension** - Chat moderation
- [ ] **Telegram bot** - Channel moderation
- [ ] **Matrix bridge** - Decentralized chat moderation
### Mobile
- [ ] **React Native module** - Native mobile support
- [ ] **Flutter plugin** - Cross-platform mobile
- [ ] **Swift package** - iOS native
- [ ] **Kotlin library** - Android native
---
## 2027 and Beyond
### AI Model Training
- [ ] **Custom model training** - Train on your organization's data
- [ ] **Federated learning** - Learn from usage without data collection
- [ ] **Transfer learning** - Fine-tune for specific domains
### Advanced AI Features
- [ ] **Agentic moderation** - AI agents that handle entire moderation workflows
- [ ] **Explainable AI** - Detailed reasoning for each decision
- [ ] **Adversarial robustness** - Resistance to prompt injection in moderation
### Community
- [ ] **Word list contributions** - Community-driven dictionary updates
- [ ] **Plugin marketplace** - Third-party extensions
- [ ] **Bounty program** - Rewards for evasion technique reports
---
## Monetization Strategy
### Open Core Model
| Tier | Price | Features |
|------|-------|----------|
| **Free (npm)** | $0 | Dictionary detection, 24 languages, AI integrations |
| **Pro API** | $9/mo | Hosted API, OCR, Audio, Higher rate limits |
| **Enterprise** | Custom | Self-hosted, SLA, Support, Custom training |
### Revenue Streams
1. **Hosted API** - Deploy on Vercel/Railway, charge per request
2. **Hugging Face Inference Endpoint** - Model hosting with HF billing
3. **Enterprise Licenses** - Self-hosted with support contracts
4. **Consulting** - Custom moderation solutions for large platforms
### Hugging Face Presence (Build Credibility)
- [ ] Publish model on HF Hub (free, builds awareness)
- [ ] Create Gradio Space (interactive demo)
- [ ] Offer Inference Endpoint (production API)
---
## Contributing
We welcome contributions! Priority areas:
1. **Language experts** - Help expand/improve language dictionaries
2. **ML engineers** - Improve detection models
3. **Platform experts** - Build integrations for your favorite platforms
4. **Security researchers** - Find and report evasion techniques
See [CONTRIBUTING.md](./CONTRIBUTING.md) for guidelines.
---
## Feature Requests
Have an idea? Open an issue with the `enhancement` label:
https://github.com/GLINCKER/glin-profanity/issues/new
---
## Version History
| Version | Date | Highlights |
|---------|------|------------|
| 3.2.1 | Jan 2026 | AI Integrations (OpenAI, LangChain, Vercel AI SDK, Semantic) |
| MCP 1.2.0 | Jan 2026 | Streaming, conversation memory, HTTP transport, 19 tools |
| 3.2.0 | Jan 2026 | MCP Server, dictionary fixes |
| 3.1.0 | Dec 2025 | Context-aware filtering, leetspeak detection |
| 3.0.0 | Nov 2025 | TypeScript rewrite, unified JS/Python API |
| 2.x | 2024 | Initial Python package |
| 1.x | 2023 | Original JavaScript package |
---
*Last updated: January 2026*
