Barclays chatbot

High-level architecture

Inspiration

Most chatbots in the banking sector remain vulnerable to privacy breaches while being apathetic to the mood of customers. Further, the present day chatbots serve mere informative purpose in that they are largely focused at reducing the workload of customer care centers.

What it does

This chatbot is capable of making real transactions on behalf of the user by accessing the core banking API directly and securely. Through our API, users can make payments, donations, and create accounts directly and securely from the chatbot interface.

The Barclays chatbot provides end-to-end encrypted message exchanges between the user and the bot. While the bot is capable of responding to normal customer inquiries (e.g. online account registration, finding nearest ATM, etc.), it additionally leverages sentiment analysis to figure out irritated customers and send messages that aer tailored to the intent of their request.

How we built it

We leverage the MTProto protocol of the Telegram messenger to provide a highly secured messaging medium. While we use DialogFlow v2 API for mapping a query's intent with plausible action, we employ Barclay's open-banking API for providing real-world testing environments. Furthermore, we use the state-of-the-art pretrained Flair model to understand the sentiment of the query and in turn deliver empathetic responses to the irritated user. For access users accounts and transactions, we rely on ThoughtMachine's core banking software, creating a seamless text interface between a chatbot and the core banking API.

Challenges we ran into

Integration between multiple technology platforms.
Most open-source sentiment classifiers respond poorly to ambiguous lexicons.

Accomplishments that we're proud of

A fully functional chatbot that allows the user to carry out all key transactions via a chatbot, accessing customer details via ThoughtMachine's cutting edge core banking stack.

What we learned

The integration of such diverse tech stacks takes whole lot of time than estimated. Also, it's you should never train your own sentiment classifier / ML model in a hackathon!