Designed and developed a financial lending application using IBM Watson conversational AI. The chatbot automated an interview and collected each applicants' financial history. The app leveraged machine learning to determine each applicant's payback risk level. Producing a $40 per applicant savings on 4,000 monthly submissions.
UX, LLC dba Artificial Savant is a technology consulting firm located in Tucson Arizona. Focused on small to medium-sized company software solutions relating to Artificial Intelligence. Artificial Savant designs and develops its own software solutions with a goal to distribute or resell the license.
The traditional loan application has evolved from paper to digital, but still remains fixed as a 'one size fits all' format. Each loan applicant is different from one another due to their unique circumstances relating to employment earnings and credit history. The current process involves two steps. The first involves the applicant completing the loan application. The second step involves the loan officer asking follow up questions about the applicant’s answers to probe for more detail and insights to better assess the risk level.
Lenders also want to minimize a loan officer’s time spent working with unqualified applicants. Lender leadership discovered that inexperienced loan officers would miss asking important questions that would have signaled the applicant was unqualified. In addition to spending unnecessary time with unqualified applicants, they would spend company resources on credit history queries that would cost $40 per applicant.
The goal is to change the fixed format to a dynamic conversational format to serve two primary purposes. The first purpose is to gain greater financial insight from each applicant while minimizing the unfavorable experience long forms tend to produce. The second purpose enables lenders to interview each applicant with consistency while reducing the interview time spent by loan officers to obtain greater financial insights.
Designed and developed a dynamic conversational home loan application that operates on IBM's Cloud platform. The application leverages IBM Watson's cognitive computing products Watson Assistant and Watson Machine Learning. The application interview can be accessed through a user interface on a web platform chatbot or phone texting application. The IBM Watson Assistant interviews the applicant as if they were chatting with a loan officer. The app dynamically adjusts based on the applicant’s answers. The answers are recorded and stored as context variables and run through IBM Watson’s machine learning model. The results are then organized in a table in ascending order from the least risk to high-risk prospective clients.
The Current Form
Universal 1003 Mortgage Application Form
Mapping the universal 1003 Mortgage Application Form into a conversational format
Developed a dialog that follows the ontological mapping to assure the necessary variables are collected credit analysis. This forces you to think about the dialog flow and assess the logical path. Each square node represents a context variable that is collected during the conversation and later used as an input feature for a machine learning model. Once the flowchart was complete, I created a CVS file to clearly list each context variable of interest. The hand-drawn flow chart map-enabled myself, developers, and potential stakeholders to see the entire conversation from a 30,000 foot view.
Convert to IBM Watson Chatbot
IBM Watson Assistant is the chat handle. Following the conversational map simplified the building process. However, building the chatbot also exposed choppiness in the conversational flow. This helped me to revise and question each conversational dialog's existence. It made me question the effectiveness and efficiency, as well as the user's experience.
Context Variables Collected for Chatbot
Each reply from the user is collected and stored as 'context variables' until the conversation completes. The context variables are pulled from the conversation in a JSON format and then converted into a CSV file. This list is now known as machine learning 'Features'. These Features need to be processed prior to ingesting into a Machine Learning model for assessment.
To avoid 'The Curse of Dimensionality' I needed to reduce dimensionality down to the core features for assessment. The thought process took time, but eventually, logic and practicing feature engineering began to improve the process. I soon discovered more opportunities to identify trends as I would leverage multiple machine learning models within a single assessment. My first model used a smaller list of features than my second model. The results exposed insights that improved prediction for future model development.
Machine Learning Model
I used a machine learning model to assess each feature (aka context variables) to produce a probable outcome of the likelihood of repaying the home loan. As the conversation captured more applicant scenarios, I revised the machine learning model to fine-tune the outcome to enable a higher level of accuracy.
Designing User Interfaces
Until now, all design and development has been on the concept, data management, and conversational structure. There are multiple user interfaces that need to be designed. The first interface is designed for the loan applicant, which is in the form of a chatbot. This will soon be expanded to voice technology devices such as Amazon Alexa and Google Assistant. I designed the chatbot to operate on mobile-first and desktop as secondary.
The second user interface is designed for financial lending companies and their loan officers. I designed the lending company's interface primarily for desktop use but enabled a responsive mobile design. This design was based on displaying a centralized dashboard for the overall information system. Loan officers and management could review application volume and data analytics, as well as drill down on the individual's application level.