Atara

Inspiration

One of the main challenges I've heard about the Puerto Rican Healthcare is the brain drain that we see with health professionals, often because of better economic conditions for them in the US. That's why I wanted to create software that would make it more economically attractive for these professionals to operate in Puerto Rico.

Then, when thinking how AI could help, voice technology came to mind, as I have been personally very impressed how far it has come.

Lastly, pretty much every clinic has an EHR, but most are hard or expensive to connect through an API, and an AI agent can't be useful without access to that data. Thus, this felt like the right opportunity to explore the challenge of wether Computer Use technology can be used mid-call, considering the latency requirements of a real-time voice call, and the complexity of these EHR UIs.

What it does

It is an AI voice agent that answer calls while driving an EHR (Electronic Health Record) software system live mid-call. This allows the agent to answer basic questions about patient's health record (status of an order, status of an appointment, etc...) and/or make changes to the patient's record as needed.

How we built it

We used ElevenLabs API to create Voice Agents, and created a custom tool that uses BrowserBase to drive a real browser mid-session. BrowserBase provides the technology to record and stream live browser sessions in the cloud.

Lastly we built a debugger interface to see the browser, the transcript, and a visualizer of the voice all real-time.

Challenges we ran into

Out-of-the-box voice agents were very verbose in their responses, and for the computer use part to work, we needed bigger models (GLM nor Gemini Flash worked; we ultimately settled in Sonnet 5.5).

Accomplishments that we're proud of

We used WebGL to create a sphere-like vizualiation of the agent talking. I have been meaning to explore this technology to build 3-D visuals in the browser.

What we learned

There are good Puerto Rican voices already in ElevenLabs, and they keep on improving their models. They will only get better from here.

What's next for Atara

We purposedly focused on appointment scheduling for this demo. Next steps would include having it interact with the rest of the parts of the EHR, and modularizing the work so as to be able to support many customers at the same time.