LUMA Project Won an Award at the AgentHack

Today, I’d like to tell you about our project LUMA — developed together with my teammates at Turkish Technology (Faik Enes Albayrak, İrem Girişmen Motorcu, Özlem Meryem Özelmas, and me) — which earned us the “Best Cross-Platform Integration” award at the UiPath AgentHack 2025 competition.

First of all, participating in the first hackathon of the Agentic AI era was a great experience. More than 400 projects from over 50 countries joined the competition alongside us. That alone shows the scale of the event and the value of the award we received. Before getting into the project details, I’d like to thank all my teammates once again for the day-and-night effort throughout the competition, and also thank UiPath again for organizing such a great event.

Why Did We Need a Project Like This?

We all share a common problem in working life: inefficient meetings. Meetings everyone complains about but can never reduce — and whose number keeps increasing day by day. According to statistics, only 37% of meetings end on time and with high-quality decisions. 80% of managers complain that their time isn’t going to their top priorities.

In a hybrid work setup, things get even more complicated. Manual note-taking is hard, information gets lost, the same topics are discussed repeatedly, yet no real progress is made. We wanted to solve this problem with an AI agent that doesn’t just take notes, but actually gets work done.

What Is LUMA? (Lightweight Unified Meeting Agent)

LUMA isn’t an ordinary meeting summarizer; it’s an agentic (agent-based) approach that can take action based on meeting outcomes.

Our core idea is this: When the meeting ends, it shouldn’t mean everything ends — real work should start then. We wanted LUMA to combine multiple AI models (OpenAI GPT-4o, Google Gemini, DeepSeek) with UiPath’s powerful automation capabilities. The diagram below is essentially the best visual summary of LUMA.

How Does LUMA Work?

I can summarize LUMA’s architecture as follows:

Dispatcher Phase

This is the starting point of the process. The workflow begins by obtaining the meeting recording. Our UiPath robot retrieves the recording and uses Google’s NotebookLM technology to generate a detailed meeting summary and extract potential action items.

Action Classifier Agent Phase

This is where the magic starts. Our Action Classifier Agent takes the extracted notes as input and analyzes them through a collaboration between UiPath Maestro and Google Gemini. The AI decides what each item is about:

• Does it require a code review? → DeepSeek steps in.
• Should a new meeting be scheduled? → The Calendar API runs.
• Should a ticket be created in Jira? → Ticket management begins.

Human-in-the-Loop Phase

We trust AI, but we don’t let go of control. Critical actions are presented to the user for approval via UiPath Action Center. When you click “Approve,” the agent goes and completes the task on your behalf.

A Real-Life Example: The Code Review Flow

Let’s say a meeting includes: “The code for project X needs to be reviewed.”

LUMA:

• Understands this task.
• Finds the related task in Jira.
• Downloads the code file (XAML).
• Sends the code to the DeepSeek model for analysis.
• Writes the result as a comment on the Jira task and updates its status.

It does all of this with no human intervention — simply because it was discussed in the meeting!

When we joined the AgentHack competition, our biggest motivation was to learn Agentic AI technology better and build a usable use case. But as the process progressed, we ended up creating a strong project — and also receiving a great award. Hope to see you again in future competitions…