meetmind

Inspiration

We started thinking about how useful it would be to have live fact-checking during things like political debates, where it’s often hard for the audience to tell if a speaker’s point actually lines up with reality. We realized we could turn that idea into a real product by applying it to a company's internal knowledge. The goal is to make meeting summaries actually reliable by catching false info or misremembered details before they turn into "facts" for the rest of the team.

What it does

Our project is a meeting bot that can join any call on any platform you use. It’s built to be flexible, so you can easily plug in your own data using MCP servers to give the bot the context it needs. Through a simple dashboard, you can set up your bots, manage who in your company can use them, and keep track of all their activity. It basically acts as a silent partner that makes sure your discussions stay grounded in what’s actually true for your business.

How we built it

⁠The agentic workflow was built using A2A's Google Agentic Developer Kit. This orchestrated the reasoning behind the fact cheking and the conversation use cases. Additionally, the bot's integration was done using recall.ai's platform for listening and responsding in real time scenarios. ⁠The communication between these 2 layers of the system, the real time and the agentic reasoning was done through Kafka queues and pub sub patterns. The main langauge we used was python pretty much across the whole projects besides the UI where we used React. ⁠A big part of our developing experience was using Google's Antigravitry IDE coupled with Anthropic's Claude Opus and Sonnet models. This way we made sure we moved quickly through the implementation making sure the product is the best experience for the user.

Challenges we ran into

Our biggest challenge was balancing the ambitious goal of real-time feedback with the absolute necessity for accuracy. In a live meeting, you can't afford to let a point slide, but providing instant context often clashes with the time needed to cross-reference deep internal knowledge bases and complex agent decision trees. We solved this by building a high-speed pipeline from the ground up, using Kafka queues and deterministic agent prompts to ensure the system is both lightning-fast and reliable. This architecture allows us to translate spoken conversation into verified facts and deliver them to the UI without any lag, giving teams the "best of both worlds" where speed never comes at the cost of truth.

Accomplishments that we're proud of

We are incredibly proud of delivering a complete, end-to-end product rather than just a collection of features; we’ve built a cohesive experience where a user can jump into a call, get real-time fact-checking, and receive a verified summary immediately. This was made possible by a scalable system design that prioritizes a clean separation of concerns, using Kafka for asynchronous communication and MCP for seamless data integration. By dividing responsibilities across the real-time, agentic, and bot layers with clear technical contracts, our team was able to work in parallel and integrate the entire architecture with minimal friction, creating a solid foundation that's ready to grow.

What we learned

On the technical side, we gained deep insights into the mechanics of multi-agent AI workflows, moving beyond simple code generation to understand how complex agents actually communicate and collaborate. Juggling various APIs, models, and frameworks forced us to stay flexible and significantly boosted our technical adaptability. At the same time, we learned a massive lesson in project management: the most effective approach is to find a clear problem and build a solid, simple solution rather than chasing perfection from day one. Pitching our concept to mentors early on was a game-changer, as their feedback helped us ruthlessly prioritize the features that mattered and cut out the noise so we could figure things out on the fly.

What's next for meetmind

We’re shifting our focus toward an on-premise solution to give our clients total control over their data. Currently, our system relies on Recall.ai, which hosts the agents and uses third-party paid models—a setup that raises valid privacy concerns for companies worried about sensitive info leaving their infrastructure. By moving the architecture in-house, we can guarantee that even when data is retrieved through an MCP, it never touches an external server where it could be exploited. To keep things user-friendly during this transition, we’re expanding our library of built-in MCPs. Beyond our current integrations with GitHub, Jira, and Google Drive, we’ll be rolling out new connectors based directly on what our clients are asking for. The goal is to keep the setup "plug-and-play" while providing the high-level security that enterprise clients expect.