Apppublic

Nishikant-dev/A11y-Agent-Pro-Max

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

App README

♿ Adaptive A11y-Env: AI for User-Centered Accessibility

Submission for Scaler Open Source Hackathon - Round 1 A Next-Generation Reinforcement Learning Environment for Web Accessibility.

Adaptive A11y-Env is an advanced Reinforcement Learning (RL) environment that trains AI agents to solve web accessibility (A11y) issues dynamically. Unlike static checkers, it implements Adaptive Web Solving—optimizing DOM structures for specific user profiles (Vision vs. Motor vs. Cognitive Impairment).

🌟 Why this project stands out (Scaler Hackathon Impact)

1.Real-World Impact: Focuses on accessibility, a critical but often overlooked area of web development.
2.Innovative Architecture: Merges RL paradigms with LLM inference, moving beyond basic prompt engineering.
3.Advanced UI/UX: Professional-grade Streamlit dashboard with real-time DOM reconstruction visualization.
4.Adaptive Contexts: Solves for specific disabilities, showing deep empathy and systems-thinking.
5.Containerized & Scalable: Ready for enterprise deployment out of the box using Docker SDK.

🚀 Key Technical Advancements

—🧬 Adaptive User Profiles: Specializes DOM reconstruction based on disability (Vision vs. Motor vs. Cognitive).
—⚖️ Dynamic Weighting: Scoring adjusts in real-time to prioritize high-impact fixes for the active profile (e.g., Image Alt-Text is weighed heavier for Vision-Impaired profiles).
—🤖 High-Fidelity Agent Integration: Connects with modern LLMs (GPT-4o, Claude 3, etc.) to perform multi-step DOM modifications natively.
—📊 Premium Analytics HUD: A state-of-the-art dashboard featuring live neural mapping and health progression tracking.

🏗️ Architecture

mermaid

graph TD
    A[Streamlit Dashboard] -->|Initializes Task| B(FastAPI RL Env)
    A -->|Observation Data| C{LLM Fix Engine}
    C -->|WCAG Violations + HTML| D[Base URL / OpenAI Proxy]
    D -->|Generated DOM Commands| C
    C -->|Action Step| B
    B -->|Reward & New Observation| A
    %% Styling
    style A fill:#2e2f38,color:#fff,stroke:#8b5cf6,stroke-width:2px
    style B fill:#2e2f38,color:#fff,stroke:#06b6d4,stroke-width:2px
    style C fill:#1a1b21,color:#facc15,stroke:#facc15,stroke-width:2px

🛠️ Tech Stack

—Frontend: Streamlit / Plotly / Lottie (Complex Animation & Visualization)
—Backend / RL Engine: FastAPI / Uvicorn (Agent Engine Core)
—DOM Logic: Python / BeautifulSoup4 (DOM Manipulation Simulation)
—AI Brain: OpenAI API / Hugging Face Inference Proxies
—DevOps: Docker / Nginx

💻 Local Setup & Development

1. Installation

Install all required dependencies:

bash

pip install -r requirements.txt

2. Running Locally

Launch the integrated backend and frontend processes:

bash

# Launch the processes in parallel
./start.sh

Alternatively, open two terminals:

—Terminal 1: uvicorn app:app --port 8000
—Terminal 2: streamlit run streamlit_app.py --server.port 8501

Access the dashboard at http://localhost:8501.

🛰️ Deployment (Hugging Face Spaces)

This project is fully structured and verified for Hugging Face Spaces using the Docker SDK.

1.Create a new Space on Hugging Face.
2.Select the Docker SDK.
3.Run the configuration script to prepare your git repository logic:

bash

   python deploy_to_hf.py

1.Push your code:

bash

   git remote add space https://huggingface.co/spaces/[YOUR_USERNAME]/[YOUR_SPACE_NAME]
   git push --force space main

🏆 Hackathon Alignment

—Environment Name: a11y-env
—Goal: Achieve a 1.0 (100%) WCAG Health Score using the fewest possible agent steps.
—Author: Nishikant Kumar (TheNishi / Nishikant-dev)
—Target: Round 1 Selection → Bangalore Finals

License

MIT License - Developed with ❤️ for making the web accessible to everyone.