Neutral Net

Real-Time Language Bias Detection

Neutral Net is an AI-powered editor designed to detect, categorize and neutralize unconscious bias in professional writing. It is built for the Lady Ada Lovelace Challenge organized by the Programming Club at IIT Kanpur.

Features

• Real-Time Multi-Model Detection - Analyzes text instantly as you type, powered by specialized AI models (Transformer-based classification, Zero-Shot NER, and Neural Coreference Resolution)

• Absolute Privacy - No part of the user's input is ever stored in permanent memory

• Bias Categorization - Detects contextual bias across four categories:

  • Stereotypes - Blocks harmful generalizations at sentence/phrase level.
  • Gendered-Terms - Flags outdated roles ("chairman"), while ignoring safe contexts.
  • Agentic/Communal Tone Skew - Detects subtle, subconscious phrasing that diminishes technical roles or amplifies hostility.
  • Pronoun Bias - Maps pronouns back to original subjects to detect forced gender roles.

• "Fix All" and Interactive Resolution - Offers one-click neutralization of all fixable biases. Users can also click individual highlights for AI-generated synonyms, custom replacements or to "Ignore" a flag.

• Document Support - Parses and extracts text from uploaded PDFs (.pdf) and Word Documents (.docx) for bulk analysis.

• Algorithmic Inclusivity Scoring - Calculates a bias score using an Exponential Decay algorithm and severity weights, ensuring short sentences aren't unfairly penalized.

• Live Analytics Dashboard - Features a Radar Chart that visually maps the distribution of biases.

• Export to TXT - Instantly download the neutralized text as .txt with a single click.

• Sub-Document LRU Caching - An optimized backend infrastructure that caches inference result and reduces latency by over 80%.

Demo and Usage

• Demo Video: Google Drive
• Live Website: neutral-net.vercel.app or Alternate Link 1, Alternate Link 2 (If none of the links open, try using mobile hotspot. IITK wifi may block them)
• Backend API: Hugging Face Space

(Note: The API may take a while to boot up from sleep on the first request).

Local Setup Instructions

Follow these steps to run Neutral Net on your local machine.

1. Clone the Repository

git clone https://github.com/harsshg31085/Neutral-Net
cd backend

2. Create and activate a virtual environment

# Windows
py -3.11 -m venv venv
.\venv\Scripts\Activate

# Mac/Linux
python3.11 -m venv venv
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Start the development server

In backend:

python manage.py runserver

In a new terminal, navigate to frontend:

cd frontend
npx http-server -p 3000

Tech Stack

• Frontend: Javascript, HTML5, CSS3 (Hosted on Vercel)

• Backend: Python 3.11, Django, Django REST Frameworks (Containerized on Huggingface Spaces)

• AI and NLP Engine:

  • Transformers: Contextual sequence classification, reasoning and rewrites.
  • GLiNER: Generalist and Lightweight Named Entity Recognition
  • fastcoref: Fast neural coreference resolution
  • spaCy: Grammatical dependency parsing

Documentation

For a much more comprehensive coverage of Neutral Net's architecture and inner workings, check out Documentation Folder:

• System Architecture
• Backend
• Frontend
• API Usage