Scaling Conversational AI in the Pre-Transformer Era

Graph with "Thousands of incoming calls per hour: high-volume / high-performance transactional and AI system working together". Details have been removed.

Google, Facebook, IBM provide NLP services using their inmense dataset. However, Enterprises have large volume and high quality domain specific data with scarce availability of training data.

Ydilo developed a new AI engine combining unsupervised dialogue multi-intent detection and other Conversational AI strategies, including NLP and ML classification.

User: I want to talk to a human agent.

System: Ok. In order to best serve you with a qualified agent, please describe your request.

User: I'm calling from my office and I have a problem with my internet connection. It is not working since this morning

System: Your incidence is with your mobile data or wifi?

User: Wifi... the wifi is not connecting to internet

System: We have reconfigured your internet gateway. Your problem should have been fixed now. Please, check again your.... #

Ydilo has been awarded "Best IT and auto-service project in 2020" 

AI in realtime conversations at scale requires cloud and software infrastructure designed to serve production and the multi-skilled (AI, NLP, content, data analytics...) teams working behind 

A high performance system for Telcos, Banks, Utilities and Enterprises requires resolving other challenges when getting this in production. In this context volume of data scales to millions of events and entities with multiple data sources and transactional systems used in real-time conversations.

Hundred of thousands inbound calls with the cloud service using AI for automatic transcription and classification. All with real-time interation with on-premise and customer cloud API transactional services.

In order to deploy daily/weekly to production a CI/CD process is required. This is integrated with an AI ML pipeline to update, train, validate and deploy versions continuously.

It is critical the operation stack with custom tools specifically developed, both in order to have a complete view of technical & business performance, from bird-eye to fine detail view, as well as new workflows to manage AI model and processes life cycle.

Note: the human with a shovel here "represents" the requirement to do this at high-availability and reliability level with multi-disciplinary internal and customer teams (e.g.: from technical -dev, qa, ops, business analyst..- to UX and linguistic experts).

In case of Ydilo we designed and developed a set of pipeline, automation and suitable tools as complete continuous learning an delivery AI + transactional production platform... a little more complex than a human with a shovel ;)