Deana.AI® is a virtual assistant that uses GTP-3 and a custom NLP model behind the scenes to understand and execute user commands. Users communicate with Deana.AI® using simple text by email and in messengers and web chatbots.
Email is a traditional way of digital communication between people in business and everyday life. Services like Gmail and Outlook help us exchange sensitive information and plan activities in advance.
Our idea was to use email to interact with artificial intelligence (AI) and give it commands. Email acts as a communication interface between the user and AI.
For example, you send the AI an email, “lunch with Dave at 5 pm,” to have your schedule updated and the invitation sent to Dave. After business lunch with Dave, you email the AI a photo of the restaurant receipt. Then the system automatically recognises the items and amounts on the bill and adds this information to your online expense report. In the end, sending these text commands to the AI takes just 5-10 seconds, saving you much time as you don’t have to update your schedule and report yourself.
Here are a few more use cases for our email-first AI assistant:
- Plan events and tasks
- Track expenses, bills, invoices, and more
- Manage files stored locally and in the cloud
- Find necessary information on the web (white papers, articles, news)
- Summarise and paraphrase documents
- Find routes and locations on the map
As Dev iNN already had a sufficient experience in AI development, we quickly formed a team that included these roles:
- Project manager
- Backend AI developer
- Frontend developer
- Data scientist
- Web designer
- QA engineer
We chose between several NLP models to find the one best fitting our technical expertise and project requirements. After some experiments, we decided on GPT-3 (Generative Pre-trained Transformer 3), which demonstrated the best results in our tests. We programmed in Python and Node.js. We fine-tuned the GPT-3 model in our knowledge domain.
As a result, we created a custom NLP model that ran on Nvidia’s GeForce 3000 series GPU. Our model required both automatic and manual correction. Humans analysed the answers generated by the model, found errors, and corrected them to improve the model further. In this way, we minimised errors and improved the model’s functioning.