This is a write-up covering the development of my startup, Teller, and the launch of our first pilot program with Brooklyn Cooperative.
Although I don’t love the name ‘chatbot’ it is the easiest way to refer to a conversational messaging agent. There are several different levels of chatbots, from simple fraud alert messages where you respond with a ‘yes’ or ‘no’ to sophisticated agents that can hold entire conversations. I imagined that the service I am building (Teller) falls somewhere in the middle, and aimed to just do one thing very well: help banks answer customer questions automatically.
I started building a chatbot back in January with the goal of developing something like a WebMD for personal finance questions. My vision was to create a database of well-researched answers and advice so that anyone could ask a question about saving, budgeting, credit, etc. and get an immediate response.
At that time, I ran a Google survey to see where people got help with personal finance questions. The top two results were as expected: Google, and friends and family. I figured both of those sources worked well enough, but could potentially be time-consuming (Google) or inaccurate (friends). I started putting together a database of questions, definitions, and explanations for basic personal finance topics and made them accessible via a text-messaging interface. My fiancé helped me come up with the name Teller.
In the bot, I’m running a Node.js server that connects the Twilio text messaging API to API.ai for natural language understanding. I store messaging data and history in a MondoDB database. I wrote the code for the Node.js server myself in order to perform backend calculations, send picture and infographic messages, and use other custom functions before responding to a user. I chose to write most of the code myself and run my…