AgentGrove

What Inspired Us

All our team members are juniors who have previously interned with the company. During our internships, we realized how hard it can be to adapt to the corporate culture — from figuring out internal tools and communication systems to understanding all those acronyms and complex project terms. It made us think about how confusing it must be for newcomers trying to navigate it all. That’s what got us interested in the NVIDIA prompt, because we wanted to build something that helps students or new employees easily learn and understand company tools and processes. While exploring, we also looked at the PNC prompt and found a similar problem that product managers face every day — managing multiple systems, tools, and complex data all at once. Out of curiosity, we reached out to product managers and even spoke to sponsors like PNC and NVIDIA to get their perspectives. After hearing about their challenges, we felt motivated to build something that could actually make an impact. That’s when we decided to create a multi-agent chatbot that helps product managers work more efficiently by connecting all their tools, simplifying daily workflows, and improving how they access information.

What We Learned

When we first started brainstorming, we came up with the idea of building a multi-agent model — and honestly, we were super excited. None of us had ever worked with AI agents before, so we knew we were going to learn a lot along the way. Our learning started with exploring what AI agents actually are, how they work, and what kind of features they could handle. We had discussions about which features to include, like summarizing PDFs, parsing slides, or analyzing images, and then divided up the research to find the right model for each part. We learned a lot about how different APIs and models communicate, how prompts affect responses, and how to test integrations properly. It also taught us teamwork — how dividing tasks, asking questions, and learning from each other can make a huge difference.

How We Built It

We started small — just brainstorming and writing down the features we wanted. Then, we researched which models and APIs could handle them best. For example, we used one model for text summarization, another for handling PDFs, and a different one for image understanding. After that, we tested everything individually before combining it. We integrated Google Slides API, Notion API, GitHub API, Nemotron API, and OpenRouter API, all working together under our multi-agent system. Each agent used a specific NVIDIA Nemotron model based on its role — the Update Summarizer Agent used Nemotron-Mini-4B-Instruct, the Document (PDF) Parser Agent used Nemotron-Nano-12B-v2-vl, the Image Parser Agent used Nemotron-Parse, and the GitHub Analyzer Agent leveraged Nemotron-Nano-9B-v2 for discussion and collaboration insights. The backend was built using Python (Flask) to handle the logic and communication between agents, while the frontend was developed using React.js for an interactive interface. Managing multiple APIs and large models at once was both exciting and challenging — especially synchronizing the agents’ responses in real time. Even though the project leaned heavily toward backend development, it gave us a deep understanding of how multi-agent systems operate, how data flows across APIs, and how models can collaborate effectively to solve complex problems.

Challenges We Faced

One of the toughest challenges was making sure all the agents worked together. Each one worked perfectly fine on its own, but once we asked them to collaborate, things got messy. We spent almost three hours debugging, testing, and adjusting how they communicated before finally getting everything to sync properly. Another challenge was balancing time between backend and frontend. Because the backend required most of our focus, the frontend ended up being lighter. Still, this entire process pushed us way out of our comfort zone. We had to work with technologies we had never touched before, learn new APIs, and stay patient when things didn’t work. In the end, it taught us the importance of communication, persistence, and stepping up to challenges — no matter how unfamiliar they seem at first.