The Debuggers

This repository serves as the dedicated space for our contributions to the Transfinitte Hackathon, where we've undertaken the project as part of the sponsored challenge presented by Netcon.

The Problem Statement by Netcon

Create a cutting-edge NLP-based AI virtual assistant to interact with users. The goal is to develop a system that not only understands and responds to user queries but also provides a personalized experience. Incorporate voice recognition technology to enhance the assistant's ability to process spoken commands and questions.

Demo Video, Presentation and ML Model:

Drive Link - https://drive.google.com/drive/folders/1G2nFJJZfAxYumeBHswDUNPAl3IRBWSi1?usp=sharing

What it does

• Performs real-time tasks as a user-friendly virtual assistant.
• Understands and interprets user queries in various languages, including English, Tamil, and Hindi, providing accurate responses.
• Implements a feedback mechanism to gather user input for continuous enhancement of performance.

Architecture

Installation Guide

Client

In the client directory, you can run:

Install dependencies:

npm install

To Start Server:

npm run dev

To Visit App:

localhost:3001

Server

In the Server directory, you can run:

Set up a virtual environment:

virtualenv -p python3 env
source ./venv/bin/activate

Install dependencies:

pip install --upgrade pip
pip install -r requirements.txt
pip install farm-haystack[colab,inference]

Download ML Model

Visit the given drive link and download the dbert.pkl and move it to /server/app/nlp_engine/pickle/dbert.pkl

To Start Server:

python main.py

How we built it

TheDebuggers_NC1 is a 3-tier application built with Flask, React, MongoDB, and Haystack pipeline. the following technologies were used for the application:

• React (Client-side Library)
• Flask (webserver/socket)
• MongoDB (database)
• Haystack (model)

Model

• The model used is DistilBERT-base-uncased-distilled-squad which belongs to a class of transformers which works on BERT Architecture, available for open use on Hugging Face.

• It has 40% less parameters than bert-base-uncased, runs 60% faster while preserving 97% of BERT's performances.

• DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT thus seems like an interesting option to put large-scaled trained Transformer model into production.

• Data Collection has been done using selenium and Beautiful Soup.Wrangled and pre-processed data contains FAQs about NetCon which has been used to Fine-Tune our NLP engine.

• F1 Score of the fine-tuned model is 86.9.

Challenges we ran into

• Time constraints
• Choosing the right model
• Web scraping netcontechnology.com, filtering, preprocessing, and gaining valuable insights.
• Incorporating feedback mechanism

Accomplishments that we're proud of

• Developed a fully functional, end-to-end chatbot application with an elegant user interface.
• Extracted data from the Netcon website to create a dataset containing 1080 question-answer pairs.
• Enabled both text and voice input options during interactions with the chatbot.
• Provided support for multiple languages for a more versatile user experience.

What we learned

• Teamwork
• Idea pooling
• NLP
• Creativity
• Problem-Solving
• Feedback and Adaptation

Future Goals

• Working on improving the accuracy of the model even futher, and optimizing the run time.
• Advanced Features: Explore the integration of advanced features such as sentiment analysis, personalized recommendations, or natural language generation for more dynamic interactions.