Cerebra AI

• 

  Cerebra AI

Inspiration

So we're reading through this challenge, and honestly, we got a bit frustrated. Every AI tool out there acts like it's this one genius that knows everything. But that's... not how life works? Like, when I bought my car last year, I didn't just walk into a dealership and trust the first salesperson. I talked to my mechanic buddy, read reviews online, asked my dad (who's weirdly good at negotiating), and then made a call. Multiple perspectives, you know? That's basically what bugged us enough to build Cerebra AI. What if instead of one chatbot pretending to be an expert on everything, we had a bunch of specialists who could actually argue with each other and figure stuff out together? Kind of like how our own brains work - different parts handling different things, but somehow it all comes together.

What it does

Cerebra AI runs like a team of specialists - we've got Finance, Risk, Shopping, and a Coordinator agent all working behind the scenes of whatever app you're already using. Each one grabs real data when they need it, then they all get into these structured arguments with each other before giving you an answer. Your normal banking or shopping app keeps doing its thing. Meanwhile, our team of agents work behind the scenes, grab data when they need it, have their discussions, and then the Coordinator packages everything into one clean answer for you. Honestly, it feels way less like talking to some generic chatbot and way more like you've got your own personal advisory board having actual debates about your situation.

How we built it

Alright, so here's how we actually built this: Everything lives on Google Kubernetes Engine because we needed it to not crash when things got busy. Each AI agent is basically its own little program - think of them as coworkers who can do their own thing but still talk to each other. We're using Google Gemini to power the agents, with their Agent Development Kit to make each one act differently. Our finance guy doesn't think like our risk assessment person, which is exactly what we wanted. The cool part is the Model Context Protocol - this lets our agents grab real information from actual apps. So if you ask about your bank balance, they're pulling from your actual account, not making stuff up. But here's where it gets interesting: Agent2Agent protocol lets them have real conversations with each other. Sometimes they disagree (which is good!), sometimes they build on each other's ideas, and occasionally they come up with solutions that surprised even us.

Challenges we ran into

Three days to build a multi-agent system? Yeah, that was ambitious. We kept wanting to add "just one more feature" until we had to literally stop ourselves. Our first version was chaos. All the agents tried talking at once - imagine a Zoom call where everyone forgot to mute. That's when we realized we needed someone to run the meeting, hence the Coordinator agent. Getting MCP working took forever. You want agents to access real data, but you also don't want them accidentally breaking your production database at 2 AM. And then there were those debugging sessions where agents would get stuck arguing about the same thing for 20 minutes, or just completely refuse to agree on anything. Turns out building AI teams has a lot in common with managing actual teams - you need good communication rules.

Accomplishments that we're proud of

Honestly, the biggest revelation was that you can bolt intelligence onto existing systems without breaking anything. MCP made that possible, and it was kind of mind-blowing. But what really caught us off guard was watching the agents develop... personality? They started solving problems in ways we never programmed them to. It was like watching a study group where everyone brings different strengths to the table. The Kubernetes deployment also taught us a lot about building things that can grow. If suddenly everyone's asking financial questions, that agent can spin up more instances without slowing down the shopping agent. We proved that you can add serious AI capabilities to existing apps without touching a single line of their original code.

What we learned

Building Cerebra AI made something clear to us: the future probably isn't one super-smart AI that does everything. It's more likely to be teams of focused AIs that can collaborate and disagree and ultimately make better decisions together. It's like how you don't use a hammer for every job, even if it's a really good hammer. Sometimes you need a screwdriver, sometimes pliers, sometimes you need to use all three at once. We saw this play out constantly with our agents. When the efficiency agent suggested one approach and the optimization agent pushed back with something better, their argument usually led to a solution neither of them would have found alone. We also gained practical skills in deploying multi-agent systems on Kubernetes, designing safe protocols for data access, and handling the complexity of communication between independent AI services.

What's next for Cerebra AI

We want to expand this thing by adding way more specialist agents for stuff like healthcare, legal advice, and education - the more expertise we can bring to the table, the better the answers get. We're also working on fixing those conversation rules so our agents stop getting stuck in endless debate loops or hitting those frustrating deadlocks. Long term, we want to see Cerebra AI working alongside apps people actually use every day, not just our demo setup. And we're curious to study whether people prefer getting answers from a team of agents versus just one chatbot voice - we want to prove that sometimes the smartest answer comes from smart people (or AIs) working together, not from one voice claiming to know it all.

Built With

• docker
• draw.io
• fastapi
• kubernetes
• postgresql
• python