GuessWhat

"Houston we have a problem!"

https://docs.google.com/presentation/d/1iOpLoi-afa1dKs8IT6wSYiSTflpyqeznw9r6D5TQxNg/edit?usp=sharing

Inspiration

Nowadays, companies use Facebook page to communicate with their customers and some how it ends up with customer feedbacks, problems and complaints and here the problems

• Too many posts on Facebook Page
• Page Admin has no time to take care of everything
• A real big problem will go “viral” if we not solve it in time especially corporate related
• What should we focus first?

Is it better to know what is the most important case to handle? Let GuessWhat Assistant to manage priority of your tasks

What it does

• Classification API
• Classify intentions from user message/conversation
• Classify sentiment from user message/conversation
• Prioritized jobs/tasks based on business rules

How we built it

Backend

• RESTful API using Flask (Python Flask)
• Training/testing dataset from Facebook (via API)
• MongoDB for a data storage
• fastText
• Wokers servers hosted on AWS

Web Frontend

• HTML
• Vue.js

Challenges we ran into

1. Post data on Facebook is unstructured, plenty of noise and we've to carefully process and clean it
2. We have to build a reliable system to repeat #1 process many times to get better models
3. Vue.js and fastText are quite new for us

Accomplishments that we're proud of

1. We can built a reliable system in just 20 hours
2. We use Vue.js and fastText for the 1st time with this project!

What we learned

1. Machine learning process like data gathering, data cleaning, model training/testing, model tuning and deployment
2. We cannot get a good model from a single ML process. We have to repeat it many times also each one takes time to get a result. So we have to had a system to support this. The system that can allow us to do the ML process many times as we want and it can give us a fast feedback.
3. Work under pressure, build a workable product in a short time frame

What's next for GuessWhat

• Some of messages like a simple question/FAQ will be escalated to an integrated Chatbot
• Some specific questions escalated to Support Specialist
• Custom business rules
• Integrated with corporate CRM

Built With

• amazon-web-services
• fasttext
• mongodb
• python
• rabbitmq
• vue.js