KOI | Devpost

Dashboard
AI Agents
Alternative Solution

Inspiration

The idea for KOI came from a personal experience. One of our teammates was stranded at an airport for three days without a clear backup plan. The most frustrating part wasn't just the delay, but having to figure out what to do next with very little certainty.

We realized that when travel plans fall apart, finding another flight is only part of the problem. There are budgets, schedules, reservations, and important commitments to consider. We wanted to build something that could help people make those decisions together instead of figuring everything out on their own.

What it does

KOI is a multi-agent travel recovery assistant that helps users explore alternatives when their transportation plans are disrupted.

It brings together four specialized agents: Travel, Budget, Commitments, and Alternatives. Each approaches the problem from a different perspective, researches possible solutions, and shares its findings with the others. The agents review each other's recommendations and work toward a recovery plan that considers the traveler's priorities.

How we built it

We built KOI using React and Vite for the frontend, FastAPI for the backend, and LangGraph with Ollama Cloud to coordinate the agents.

The agents research transportation options using web search and retrieved sources. LangGraph manages two rounds of deliberation, allowing agents to share their findings, raise concerns, and revise recommendations. We used Server-Sent Events to display their progress and conversations in the interface.

We also added structured validation and source checks to reduce unsupported recommendations.

Challenges and what we learned

One of our biggest challenges was getting multiple agents to work together meaningfully rather than simply generating four separate answers. We had to think carefully about how they shared information, handled conflicting priorities, and evaluated one another's recommendations.

We also learned how important reliability is when working with real-world travel information. Prices and availability change constantly, so we focused on making recommendations traceable and avoiding guesses when information couldn't be verified.

Exploring Cisco's CoffeeAGNTCY lab also helped us better understand agent coordination and communication.

What's next

We'd like to integrate transportation-provider APIs, introduce personalized travel profiles, and add more specialized agents, such as an Accommodation Agent. We also want to improve response times and explore AGNTCY's A2A and SLIM capabilities as the system grows.

Ultimately, we want KOI to make travel disruptions a little less overwhelming by helping people understand their options and find a practical way forward.