AI assistant for American Growth Insurance

A user asks the assistant to do a piece of work. It works through the task with its tools and shows its progress, and its answers stream in as they are written.

• Tools and workflows. A dynamic registry of tools the assistant can call, and a workflow manager.
• Documents. Users upload files in several formats, and it works from what they contain.
• Web research. Real-time search, with the sources cited.
• Isolated code execution. Code runs in a Daytona sandbox.
• Generated files. It can build a PowerPoint deck from its research, and simple web apps.

The agent runs on the Suna framework. My part was making it a feature of AGI’s app.

• The interface inside their app. A chat in AGI’s existing React app, with file upload, drag and drop, and conversation threads tied to projects.
• Live tool output. I ported the UI and the responses for each of the agent’s tools into AGI’s app, so each tool’s output streams in while the agent works.
• Their sign-in and their data. Authentication and data stay in sync between AGI’s backend (Express, PostgreSQL, AWS Cognito) and the AI side.
• Infrastructure as code. Terraform and Terragrunt, with CI/CD, monitoring and autoscaling.
• Speed under load. Caching and connection pooling.
• Audit logging. An audit trail for the assistant’s activity.

The architecture is modular, and I used AI code generation under supervision.