gg-friggin-ez

Product Introduction

1. Definition: gg-friggin-ez is a Node.js library for real-time, multilingual content moderation and profanity screening. It is a technical tool designed for developers to integrate into chat systems, social platforms, gaming environments, and comment sections to automatically detect toxic language and explicit content.
2. Core Value Proposition: It exists to solve the critical problem of scalable, accurate, and affordable real-time text moderation, specifically against sophisticated user evasion tactics like leetspeak, character spacing, and romanized profanity across multiple languages, including Indic languages like Tamil, Hindi, and Bengali.

Main Features

1. System 1 Model Integration: The library is powered by reflex-speed AI classifiers, specifically TypeSafe AI's Jev model by default. These are not generative LLMs but decision models that return calibrated probabilities for categories like toxicity and profanity in sub-500ms, enabling real-time use cases without the latency or cost of a full LLM call.
2. Evasion-Aware Detection: The screener is explicitly engineered to detect and classify common obfuscation techniques that bypass traditional keyword filters. This includes symbol/number swaps (e.g., sh1t), spaced characters (p a i t h i y a m), repeated characters, mixed-script text (e.g., chuतिya), and even 2D ASCII art drawings used to convey hostile gestures.
3. Native Multilingual & Indic Support: Unlike many English-centric moderation tools, gg-friggin-ez provides robust support for code-mixed and romanized text in South Asian languages. It can detect profanity and toxicity in transliterated Tamil (Tanglish), Hindi (Hinglish), Kannada (Kanglish), Telugu, Bengali, and Malayalam without requiring language-specific training data from the user.
4. Configurable Policy Engine: The AI model only provides classification scores (e.g., severityScore: 1.95). The library applies deterministic, user-configurable thresholds on the client side to decide actions (ALLOW, SUSPICIOUS_REVIEW, AUTO_CENSOR, AUTO_BAN). This separates the classification logic from the moderation policy, giving developers full control.
5. Pluggable Architecture: While Jev is the default, the system is designed to be model-agnostic. Developers can configure the library to use other System 1 decision models, such as a local instance of Laya by ConvAI Innovations, or any custom HTTP endpoint that returns data in the expected format, ensuring future-proofing against model changes.

Problems Solved

1. Pain Point: The failure of static keyword denylists and slow, expensive LLM APIs to catch evasive, real-world toxic content in user-generated text, especially in multilingual and code-mixed environments.
2. Target Audience: Backend Node.js developers, DevOps engineers, and product managers building social media platforms, online games with chat, comment systems, customer support portals, and community forums that require real-time content moderation at scale.
3. Use Cases: Essential for screening in-game text chat for offensive leetspeak, moderating user comments on a regional news site with code-mixed language, filtering toxic messages in a live-streaming chatroom, and automating initial triage for a human review queue in a social media app.

Unique Advantages

1. Differentiation: Compared to generic moderation APIs (e.g., from large cloud providers), gg-friggin-ez is specifically tuned for evasion patterns and Indic language romanization. Compared to open-source keyword filters, it uses an AI model for context-aware understanding. Compared to using a general-purpose LLM like GPT-4 for moderation, it is 10-20x faster and 100x cheaper per call.
2. Key Innovation: Its core innovation is the combination of a high-speed, low-cost "System 1" decision model with a client-side rule engine specifically designed for adversarial text patterns. This hybrid approach delivers high accuracy on evasive content with the latency and cost profile necessary for real-time applications.

Frequently Asked Questions (FAQ)

1. How does gg-friggin-ez handle different languages and scripts? It uses a multilingual AI classifier (Jev) trained on a diverse dataset, allowing it to understand context and toxicity in romanized forms of languages like Hindi, Tamil, and Bengali without relying on script-specific keyword lists, effectively detecting code-mixed phrases like "abe chutiya" or "nee p00da."
2. What is the cost of running gg-friggin-ez in production? Using the default Jev model via OpenRouter, the cost is approximately $0.000004 per message (or $0.042 per 1 million input tokens). There is no charge for output tokens, making it significantly cheaper than using generative LLMs for content moderation tasks.
3. Can I use gg-friggin-ez without an API key for testing? No, the npm package requires a valid API key (e.g., OPENROUTER_API_KEY) to call the Jev model. However, the project's live demo page includes a fallback heuristic mode for interface testing without a key, but production use mandates an API key for inference.
4. How fast is the gg-friggin-ez profanity screener? End-to-end latency typically ranges from 50ms to 500ms, as measured by the latencyMs field in the response. This performance is achieved by using specialized System 1 models instead of slower, conversational LLMs, making it suitable for real-time chat moderation.
5. Does gg-friggin-ez support custom moderation rules and thresholds? Yes, developers have full control. You can adjust the confidence thresholds for review, censor, and ban actions per API call. Furthermore, you can implement a completely custom schema by using the createScreener() function with your own set of classification questions for the AI model.