Machine learning digital assistant and image recognition
RS Components – Chatbot and Image Recognition
RS Components operates in over 30 countries and is the number one high service distributor of electronics and maintenance products across Europe and Asia Pacific.
Problem / Opportunity
The aim of the project was to contribute to RS Component’s long term strategic goal to be the market leader in customer experience and to create a foundation upon which disruptive new customer interfaces could be built.
We partnered our machine learning and image recognition experts with their innovation department, customer service specialists and sales team to form a comprehensive task force. The whole task force then attended a workshop in which different approaches and needs were discussed. Once a consensus was reached, initial user stories were established to help guide the development process.
We then proceeded to build the MVP chatbot framework and image recognition system, which was refined and expanded with additional user stories in feedback driven iterations. This was then released to the market in a staged rollout, with the explicit aim of getting real market feedback as early as possible.
There were two main technical challenges to the project; intelligent chat and product identification via image and unlike conventional software, each of these challenges also required a two stage development approach of training and then runtime-operation.
In order to solve the intelligent chat challenge we leveraged Dialog Flow and machine learning to create a chatbot harness which utilised annotated chat templates and chat transcripts from real user interactions with customer services. This allowed us to develop a solution which was capable of maintaining context over multiple interactions, request clarifying information from users and fulfill customer requests by leveraging upstream systems.
Product identification via image is a hard problem to solve, especially when trying to classify user generated images taken under random conditions. In order to achieve greater accuracy of classification, we decided to enlarge the training data set of images. This was achieved by taking the perfect catalogue images and applying transformations which mirrored the most common photo errors, such as being out of focus, unwanted shadows, distortion etc. This transformed data set was then added to the original data set and formed the basis for training the image recognition neural nets.
Increased accuracy and capability of both the intelligent chat and image recognition system will also be achieved through a self-learning system, that utilises each interaction to expand its training set. Allowing the system to become smarter and more resilient as customer use increases.
The rollout of the chatbot aims to increase customer satisfaction through faster response times, 24/7 availability, lower overall support costs and much improved product identification methods compared to human agents.