Tchiko

Inspiration

The inspiration for this project came from visiting Tchibo stores in Germany. As a coffee lover, I'm always up for a cup of coffee and so is Tchiko. He is not just any chatbot - he wants to believe that his name came from Tchibo and coffee and is ready to help anyone!

What it does

Tchiko starts his morning from a cup of coffee and is ready to engage with customers. He has access to Tchibo articles and products and has voraciously read everything about them. Tchiko can answer customers requests, filter the database and help with the potential order. He can handle multiple items in a single order and inform customers of the final order price. Tchiko speaks German, but he is already dreaming of acquiring a few other languages (Czech? Hungarian?) once he has finished translating articles from the database.

How we built it

Tchiko uses recent advances in the fields of natural language understanding and machine learning, and is built using Rasa, an open-source framework for conversational AI, sentence encoders from TensorFlow to assess sentence similarity, and underlying spacy library for tokenization, entity classification, and much more. One can test it in Rasa X, from the command line, but we have also added Flask API. The first models have been for entities/items that can be found in orders.

Challenges we ran into

Tchibo has a product taxonomy of several levels, and I have used some of those, but it would be great to talk to business to identify which depth of taxonomy may be the most relevant for consumer. Also, the business input is invaluable when it comes to the customer journey; the first version of Tchiko is based on a couple of assumptions about customer intents (customer knows what she is searching for, customer needs help to order, order confirmation).

Accomplishments that we're proud of

I am happy to be using the state-of-the art technology in chatbot development to ensure a chatbot can learn new entities and recognise them in various contexts. It was great to fetch data from Tchibo API to test Tchiko on some real stuff, including images, prices, gender, etc. It has been a sprint (started only recently), but am proud of what has become possible in a few days.

What we learned

My background is in ML and NLP, however I have never built a chatbot before. In this hackathon, I have learnt quite a bit, from building chatbots from scratch, using Rasa, to envisioning a customer journey for helpers like Tchiko. Loved the experience, danke!

What's next for Tchiko

Tchiko has just started! He has big plans for multi-lingual support and seamless integration across various channels and touch points. While he cannot offer to smell coffee he's recommending, he would love to direct customers to a nearest store, should they be interested in it. Tchiko is using product taxonomy and would love to guide customers through it. Whenever Tchiko converses with customers he learns more about their intents and he would love to improve further. Also, he would like to finally start talking and the next step would be to use text-to-speech APIs to accomplish it.

Built With

• flask
• google-cloud
• python
• rasa
• spacy
• tchiboapi