About the Project

This project was inspired by a common problem in professional work: meetings generate decisions, commitments, and follow-ups, but those action items are often lost in notes, chat threads, or after-the-fact emails. I wanted to build an agent that quietly handles the boring but valuable parts of work — extracting action items, assigning owners, tracking deadlines, and only interrupting the user when something truly needs attention. That made it a strong fit for the AWS “Agents for Humans” hackathon, especially in the Professional Agents track.

I learned a lot about building agentic systems with the Strands SDK, designing tool-based workflows, and making automation feel trustworthy instead of noisy. One of the biggest lessons was that the real challenge isn’t just extracting tasks — it’s deciding when the agent should act silently and when it should escalate. I also learned a lot about deployment, model access, and IAM configuration when moving from a local demo to an AgentCore runtime.

I built the project around a Strands-based agent with separate tools for transcript extraction, task creation, email drafting, and escalation checks. The system processes a meeting transcript, creates structured tasks, stores state across sessions, and reviews open items for overdue or ambiguous ownership. I also built a clean dashboard to show meetings processed, tasks grouped by status, and an activity log that clearly distinguishes auto-handled actions from escalations.

Built With

• bedrock
• fastapi
• groq
• python
• react
• strand
• tailwind