Overview

Build a Generative Support Agent for NeuroTech that can understand visitor questions in Arabic or English and generate accurate, human-like responses — without simply fetching a pre-written answer. The system should be able to handle diverse support scenarios such as course info, subscriptions, technical issues, pricing, or general questions about the platform.

What you need to build

• Dataset: Create or collect a dataset of 1500+ user–assistant pairs in Arabic and English. Each entry should include user_query, assistant_answer, and optional intent. Use real or simulated questions from support, course info, pricing, login, or general FAQs.
• Model: Train or fine-tune a generative model (e.g. mT5, Flan-T5, AraT5, Llama-2) or build a RAG-style pipeline (retrieval + generation). The agent must understand and generate natural answers.
• Language Handling: Detect the input language automatically and respond in the same language.
• Interface: Build a simple chatbot UI (Streamlit / Gradio / FastAPI) with NeuroTech theme colors (navy/gold).
• Report & Video: 3-page report (data, model, examples, analysis) + 7–10 min video demo explaining the system.

Technical guidance (non-mandatory)

• Preprocessing: Normalize Arabic (remove diacritics, Tatweel, unify Alef/Ya/Ta-marbuta), lowercase English, remove URLs/emojis.
• Option 1: Fine-tune a multilingual generative model (mT5 / Flan-T5 / AraT5).
• Option 2: Use a RAG pipeline (retrieval + generation) with embeddings and a knowledge base (e.g. FAQ/JSON).
• Prompting: Design structured prompts like: "You are a helpful NeuroTech assistant that answers in the user's language."
• Evaluation: Measure BLEU / ROUGE / semantic similarity or qualitative examples of generated responses.

Deliverables

• Demo: Chatbot interface (Streamlit / Gradio / Web) showing user input → generated response.
• Code: End-to-end pipeline (training → inference) + requirements.txt.
• Dataset: CSV/JSON (1500+ pairs) + data_readme.md explaining sources & preprocessing.
• Report: 3 pages describing data, models, architecture, evaluation, and sample outputs.
• Video (mandatory): 7–10 min screen-recorded demo explaining the approach and showing results.

Submission Package

• Demo Link: working app / Hugging Face Space / hosted Streamlit
• Code/Repo: GitHub or ZIP file
• Dataset: uploaded CSV/JSON
• Report: 3-page PDF/DOC
• Video Link: YouTube unlisted or Google Drive (required)