Voximplant Avatar - is an NLU-powered cloud voice-and-chat-bots platform, which allows you create intelligent omni-channel agents for your contact center. It’s deeply integrated into Voximplant platform, which allows you create cloud contact center without worrying about infrastructure. My mission at this project was to build PoC, gather a team and drive the product development.
Modern communication automation relies heavily on Machine Learning & modern NLP neural networks for tasks such as intent classification & entities extraction. And recent breakthroughs in this area, like transformers architecture & transfer learning, levelled up the quality of NLP solutions. But it all have it’s own price: descent models are pretty heavy, require expensive GPU-powered infrastructure to work in real-time and needy for ML engineering expertise to set it all up.
*data from May 2022 on conversational benchmark
Luckily another trend has beed emerged in recent years: AutoML. It’s a concept of mainly cloud-based services, which takes all heavy lifting on itselves. The only thing you need to do - is just feed the service with data, and it will give you ready-to-use ML service for your needs.
In case of Voximplant Avatar platform, we decided to let our users train State-of-the-art models (based on XLM Roberta) with their own data. This gives best-in-class recognition quality. In order to have ‘on-demand-tarification’, we’ve implemented several infrastructural know-hows to manage these models efficiently on our GPU resources, and give the quality of in-house development models with price tag of ‘one-size-fits-all’ platform solution.
Serverless dialog scenario coding
While no-code solutions serves well for simple scenarios, real-world automation usually require tricky integrations & complicated business logic of conversation flow. We wanted to give the most flexible solutions, so Avatar offers a cloud IDE for writing dialog scenario in Java Script. At runtime this code is executed in serverless manner on Voximplant Platform.
Simple integrations with telephony & text channels