Inspiration

  • Originally thinking of personal coach (for gym, singing, programming, etc)
  • POC is aggregating today's information --> B2B potential for the AGI secretary, with lots of potential for GPT plug-ins to support the enterprise customer's multiple third-party integrations

What it does

Airi - Your AGI Secretary!

  • a 3D "talking head" avatar
  • performs text-to-speech and outputs it with a natural voice, supporting plug-ins like AWS Polly (default), Eleven Labs (once its text-to-speech becomes much faster...!)
  • utilizes text content powered by ChatGPT + Plugins
    • weather data
    • horoscope data
    • news data

How we built it

  • Made "plug-ins" (API endpoints) to be used by AutoGPT to shortcut the common Airi functionalities.
  • Get user profile data from MongoDB + inject it into the prompt to provide personalized context, and hence personalized responses!
  • Unity to generate 3D "talking head" avatar
  • AWS Polly (default) and Eleven Labs (tested and gave up) for text-to-speech output

Challenges we ran into

  • We did not have ChatGPT Plugins access, so used AutoGPT.
  • Issues getting JSON POST working on Unity, switched to query form variable.

Accomplishments that we're proud of

  • Get AutoGPT working with the helper functions we write
  • A suite of useful POC helper functions, which can be easily made into a single plug-in.
  • Generalizable 3D "talking head" interface which could work with any text-based input! So any kind of conversational responses, which many plug-in outputs could be converted into.

What we learned

  • The promise offered by GPT Plugins is incredibly exciting, even without Plugins access.
  • Lots come down to prompt engineering still, to have the user context properly injected + AutoGPT fulfill the request as expected.

What's next for Airi

  • Support additional roles -- additional 3D avatar and response creation patterns
    • e.g. a gym coach, eg for strength training, marathons, whatever -- would ask about the user's fitness profile + inject that into prompts for customized fitness recommendations + generate a dynamic strength training plan

Built With

Share this project:

Updates