Our client is a startup that helps businesses transform by developing corporate training programs for them.

The company effectively improves change management, enabling one-on-one interaction with every employee on the changes and reporting on possible risks and issues to managers.

With the COVID outbreak, the field of change management started gaining more attention, while the demand for these services began to increase rapidly. To address the growing demand, in February 2020, our client introduced a micro-coaching service that provides customers with answers to their questions via text or chat. The solution was warmly embraced by users and immediately received a lot of positive feedback. 

But despite the initial success, the company faced certain challenges over time. First, their chatbot had a monolithic architecture, which made it hard to add new functionality. Second, the chatbot could not recognize users’ emotions, moods, and intents and express empathy. Third, as the company's customer base continued to grow, it became harder for the chatbot to carry a heavy workload. This prompted the need for scaling.

That’s why the company came to Wetelo to achieve the following goals:

• Re-architect the chatbot to make it easier to add new functionality
• Make the chatbot more intelligent for more efficient change management
• Improve the overall maintainability and scalability of the chatbot

Product discovery to define how to enhance the virtual coach

To improve our client's solution and decide on what features to add, we started with product discovery. During this phase, we interacted with the chatbot as a user and analyzed it to better understand what functionality it lacked. We came to the following conclusions:

• It is important for users to get personalized messages with the bot mentioning their names. 
• The virtual coach has to suggest different options for users based on their replies, and then reach out with a follow-up asking about their progress. 

To address these needs, we offered to build a GPT-3 module that would enable the chatbot to come up with user-tailored messages during 1:1 conversations, recognize users’ moods and emotions, show empathy, and encourage them to embrace the changes and set goals. 

Building a sophisticated chat GPT-3 module and increasing the product’s scalability

To meet our client’s requirements, we broke down the development process into several stages, starting with architectural refactoring. 

Our first step was to break up the monolithic application into microservices. Together with our client, we made the decision to move to microservices because of their greater flexibility, better convenience, and the ability to build more manageable services that can be developed, deployed, and scaled independently. We successfully moved the product to the microservices-based architecture, which allowed us to add new features seamlessly.

Our second step was to test and compare GPT-3 models like Davinci, Ada, and Curie and choose the best option according to the client’s needs. The results were as follows:

1. Ada model produced acceptable results in terms of sentiment analysis.

2. Curie model did an excellent job with simple paraphrasing.

3. Finally, Davinci combined the functionality of both Ada and Curie, and it was also the best model in terms of understanding the text intent, which was crucial for our client’s solution. 

We decided to use the Davinci model to analyze employees’ emotions expressed in text messages.

Our third step was to build a module that would receive messages from employees and reply to them with text generated by GPT-3. With this module, the chatbot could understand employees’ moods, paraphrase employees’ messages, and express empathy. For example:

Employee: “I don’t really know what I’m doing. Sometimes it feels like I’m walking around in circles, things are just so stressful that I never get anywhere”

Chatbot: “It sounds like you’re feeling overwhelmed and you’re not sure what to do. Everything feels so stressful and it feels like you’re not making any progress”. 

Our fourth step was to conduct sentiment analysis and check whether GPT-3 can define the characteristics of emotions. For this task, our team used sentiment-roberta-large-english model and Huggiing Face API. After testing, we concluded that GPT-3 worked as intended, as it was able to distinguish negative sentiment from neutral. Based on the results provided by the chatbot, human managers could better communicate with employees, improve their morale levels and prevent risks associated with burnout and turnover. 

To deal with sudden traffic spikes and ensure product scalability, together with our client, we made a decision to migrate the solution to the cloud. That’s why our final step was to perform the product migration to AWS, as this service provider offered more opportunities for developing, maintaining, and scaling the solution, while allowing to deal with heavy traffic hassle-free.