Lots of Business Hours of Customer Support are Wasted due to Some Q&A,Doubt Resolution and Other Stuffs . So to Provide and Take Every request from the customers , Company resources are inefficiently increased and it's impact business of the company.

• To automate customer service.
• Reduce Costing and Increases Efficiency by Providing Support in More Regional Languages

MegatronBot is a fully fleged chatbot with easy to update, integrate with website, easy to deploy in any cloud services like AWS, GCP and azure with a capibility to work in production enviorment.Megatron accepts various formats of inputs you can give a text input, you can also give a Speech as a input.You can Ask in any language that is accepted by Google.

Megatron was build by the team of ineuron.ai where I have worked upon a task of Intent Classification and State Tracking.The need of the MegatronBot came as ineuron.ai has a overwhemling responses and queries asked by the student over their newly released courses, as their is a Skype Support team to clarify their queries always but they cant be available 24x7 hours due to this ineuron.ai wants to build a such a great solution where the user can clarify their queries anytime without waiting in a queue for hours.

Building a Megatron like ChatBot requires a huge amount of data and various state of the art components.So, the first task is to get a large data,The dataset was creatd by scaraping the queries and answers asked by students over a year to their Skype support team, the dataset include of 20k sentences which were transformed and added to CSV and json format.

I have tried many State of the art language models from BERT-large to DialoGPT to RoBERTa but got an awesome results over Distilled BERT Models with ELMO embeddings.The model then trained over the 20k queries after preprocessing then adding tokens like [SEP], [START], [END] and [EOS].

Here we are using Distilbert Model after testing around 22 models including BERT,ALBert,ProphetNet,Electra,Elmo,T5,Bart and Others. Please Checkout HuggingFace website for the NLP models .

Finally, Distilbert Model was Choosen based on the Performance .

We are Using Google Api's for the Web Scrapping in case DistilBert Models fails to give the Answer from the trained dataset provided by iNeuron.ai. Then The Megatron does Scrapping from google Search engine using the api's and then it's been Summarized .

In Case Customers is not satisfied with the Bot then Bot will lead to Support Agents.

In this Way it will take 40-45 % load for now and After 6 months the prediction is of about 70%.

#Step 1 :-- Configure the Enviroment settings in the IDE 
# I am using Pycharm and Conda
#Step 2 :-- Open the Terminal of Pycharm
pip install -r requirements.txt

Step 3 :-- Run the app.py
python app.py