Aura
An AI chief of staff with a face. You talk to it, it does real work in your Gmail, Calendar, GitHub, Notion, Linear and Slack.
The problem
Talking to an AI is easy. Getting one to actually act inside your accounts, with your credentials, is the hard part. Most avatar platforms want to run the AI for you, which means handing a third party your tokens.
What I built
Aura is a real-time video avatar you speak to, backed by a LangGraph agent on my own server. The avatar vendor never runs the AI: I exposed my backend as an OpenAI-compatible endpoint, so their platform believes it is talking to a model while the agent and every credential stay on my side.
The agent holds authenticated connections to Gmail, Google Calendar, GitHub, Notion, Linear and Slack, and does the work itself: sends the mail, books the meeting, files the issue.
How it went
Speed decided whether it feels alive. Nobody waits for a face to think, so the pipeline is tuned to speak its first token in 187ms at the median on the fast route, measured over a 340-turn scripted session. It is deployed and usable, with a test suite of 90+ tests behind it. The code stays private; the deployed product is the demo.
Numbers
- 187ms
- 90+
- 6