Projects / AI Voice Agents

AI VOICE agents

Grace is our voice AI agent framework. Users create and deploy their own versions of Grace by defining playbooks, attaching knowledge bases, and coaching behavior through interactive sessions.

A core design challenge was translating these abstract system concepts into interfaces that users could understand, configure, and iteratively improve.

Playbooks

Defining agent behavior

The Playbook defines an agent’s workflow, similar to a procedure a call center agent would follow.

Rather than designing a single linear setup flow, I chose to model agent behavior as a sequence of explicit steps. This made workflows inspectable and editable, allowing users to reason about agent behavior, debug issues, and update logic incrementally as requirements evolved.

Coaching Sessions

Training through conversations

Users coach their agent by speaking directly to it in test conversations. These sessions simulate real calls, allowing users to hear how the agent responds in context.

During a conversation, users can intervene and make corrections to both the information the agent uses and the actions or dialog it produces. This creates a tight feedback loop where agent behavior is shaped through example.

I intentionally designed this interaction to mirror how humans train other humans: by role-playing scenarios and giving feedback in the moment.

knowledge bases

Providing Context

Knowledge bases allow users to provide domain knowledge and context through plaintext documents or external URLs.