CallMeBot was designed to help a local British car dealer with car sales.
This calling bot was designed to call the customers, ask them questions about the cars they want to sell or buy, and then, based on the conversation results, give an offer on selling or buying a car.
CallMeBot was created for the auto industry. But such a bot can be used in any sales industry that requires working with cold calling specialists and collecting information about the products or services. In this case study, we are answering one of the most important business questions: How to make a bot for a website?
Location: USA
Industry: Transportation
Product: Chatbot with machine learning
The scope of our work: back-end, machine learning
Solutions: Twilio, CNN, Haar Cascades, heuristic filters
Business goal: a bot had to replace the cold calling specialists.
This bot needed to be communicative, understand human questions and respond in a human-like manner.
Technical goal: a bot had to check and collect the information about the clients and give them a relevant offer on selling or buying a car.
We had to create such a bot that would not only be able to understand human speech like other bots for a website, but also analyze it, and give an appropriate response.
Challenge: Speech to text/text to speech translation
We needed our bot to translate text to speech and vice versa.
Solutions:
CMB Speech is based on Google TTS (Text To Speech) or Amazon TTS, but it is possible to install the real human voice recording. It is also possible to use a keyboard for numerical data to ensure that there are no voice-recognition mistakes.
We used the Twilio platform to convert Speech to Text (STT). Once each response is recorded, it is extracted with STT (Speech To Text) and checked for the markers.
Other solutions:
We used various kinds of heuristic filters and methods, based on the Haar cascades for 95-99% of the accuracy for recognition.
We used Haar cascades for car number plates recognition. It was important to define the plates boundaries and to clear an image from the shadows or any other side information.
We applied CNN (convolutional neural network). It was needed for numbers and letters recognition.
Results:
Working on CallMeBot project, we provided our product with extended functionality.
Now our voice bot has a lot of features and is able to:
Check car ads and gather information about mileage, residing city, registration number, service history, and phone number
Give interactive responses to clients’ requests
Receive, process, and send data
Recognize license plate numbers using a picture
Parse advertising sites for information
Work on Twilio with good stability and scalability
Work with any region and any type of phone number (both mobile and landline)
Detect a client’s region and use it to make phone calls to this client
Be easily adapted to other tasks, so the solution itself is all-purpose. It is possible to change the scripts and work logic and adapt CallBot to different tasks and business when it is necessary to recognize various objects, such as faces, cars, checks, and etc.
Business results:
CallMeBot can process up to 5,000 calls daily and perform up to 50 client calls at a time
The automotive company reduced the required number of cold calling specialist
Currently, we are teaching our bot to work with different regions (by adding the client’s regional phone number) and modes (allowing the clients to call bot in case they need to)
Veeqo is an inventory and shipping platform for e-commerce. It allows managing orders from multiple channels and keeping track of inventory from multiple warehouses.
