Using the Rikai Bot Application Platform my team built a chatbot for MeadJohnson, to answer parent's questions about baby nutrition and promote sales of their baby food formula.

Given the heavy regulation around medical information, it was critical our matching algorithm would not generate false positives, so I modified the classifier to combine entity matching as another layer of signal.


Using our RiBot scripting language, Rikai's chatbot developers were able to create "user journeys" - sequences of questions and follow-up answers.

Clustering of Unlabeled Data


In this case the client gave us a huge unlabeled dump of chatlogs from their previous human support agents. We used python + gensim to create clusters of questions and answers to find the most frequent types of questions from their customers.

NLP intent / entity phrases


After clustering we further broke down the conversations by intents and entities, before preparing responses. For the top few types of questions we modeled out custom journeys with follow-up questions.