Inspiration
Being your own boss sounds great until you realize that, quite often, it means arguing with yourself at 2 a.m. You are the CEO, the salesperson, the person doing the numbers, the one building the product, and sometimes the one cleaning up the mess afterwards.
And when you work alone, or with a very small team, there is usually nobody around to challenge an idea before you spend money, time, or a ridiculous amount of tokens building it. That is where Junta Directiva AI came from.
Big companies have something small businesses and solopreneurs rarely have: several experienced people around a table, looking at the same problem from different angles.
I wanted to build a version of that for people who do not have a boardroom, a CFO, a CTO, a marketing department, or a consultant sitting next door. Not a chatbot that agrees with you. A board that argues with you.
What it does
You bring a real decision, problem, opportunity, crisis, project, negotiation, or pitch. You can also give the board real context: PDFs, Word documents, Markdown files, URLs, notes, your available budget, how much time you actually have, and whether you are working alone or with a team.
Then you choose who should sit at the table. There are up to 12 specialist directors covering strategy, finance, marketing, operations, legal, technology, sales, product, people, data and experience. And then there is Jottarina, the Chief Reality Officer.
Her job is basically to spot self-deception, overengineering, brilliant ideas that may not be quite so brilliant, and the elephant in the room everybody else is politely ignoring. The selected directors analyze the problem in parallel.
They do not simply produce twelve isolated answers. They cross-check their findings, expose disagreements, and then the Chairman takes that mess of opinions and turns it into something you can actually use:
- what to do;
- what not to do;
- under which conditions;
- what the risks are;
- what should happen next;
- and a full 30/60/90 action plan. And the meeting does not end with the verdict.
You can challenge it, add new evidence, ask what would make the board change its mind, attach more material, continue working with the Chairman, and export the result as a professional PDF. Because a recommendation without execution is just another opinion. The hackathon build is completely free for judges: no payment, no API key, no signup and no usage cap.
How I built it
The frontend is built with React + Vite and deployed on Firebase Hosting. When a user convenes the board, the frontend sends the case to a FastAPI backend running on Cloud Run. The backend creates the session, returns a session ID immediately, and the board keeps working in the background. The specialist directors are real agents built with Google ADK, using Gemini 3.5 Flash through Vertex AI. They run in bounded parallel batches rather than waiting for one another sequentially. Their progress, completed analyses, contrast round, pause/resume state and final verdict are stored in Firestore. The browser subscribes to Firestore in real time, so the user can actually watch the board work while the backend continues independently. All Gemini calls happen server-side using the Cloud Run service account. There are no model credentials exposed in the browser. That was important to me because I wanted the hackathon version to behave like a real application, not like a frontend demo calling an API directly.
Challenges I ran into
The easy part was making twelve AI personalities give opinions. The difficult part was making those opinions feel like a board. My first instinct was to run everything sequentially. It worked, but it was slow and felt like watching twelve chat responses arrive one after another. That was not what I wanted. So I changed the orchestration. The directors now work in parallel batches, then go through a focused contrast round before the Chairman produces the final synthesis. Another challenge was making the user understand that something meaningful was happening while the agents were still working.
That led to the realtime Firestore workflow: the user can see who is analyzing, what has finished, when the board is contrasting findings, and when the Chairman takes over. The other big lesson was that the verdict is not the end of a business decision. Usually it is the beginning. That is why I added the Chairman follow-up session, new evidence, attachments, reconducting the decision, and the full operational report. I did not want the user to finish with a clever paragraph. I wanted them to finish with a plan.
What I am proud of
The thing I am most proud of is that Junta Directiva AI is not one prompt pretending to be twelve experts. Each director is an actual specialized agent with a different responsibility. They work independently, their findings persist, they contrast them, and another agent synthesizes the result.I am also proud that the system respects the user's real situation. If someone has zero budget and five hours a week, the board should not recommend a $5,000 stack and a six-person marketing team. That sounds obvious, but AI tools do it all the time. This project is aimed at people who actually have to execute the recommendation themselves.
What I learned
The main thing I learned is that agentic systems need more than autonomy. They need visibility. If an AI system is going to work for several minutes, users need to understand what is happening, who is doing what, and where the result came from. The other thing I learned is that disagreement is useful. If every agent reaches exactly the same conclusion, there is probably not much point having a board. The value comes from different perspectives colliding before a decision is made. Google ADK, Vertex AI, Cloud Run and Firestore gave me a very practical way to build that workflow while keeping the human in control.
What's next
The next step is to make the board remember previous decisions and outcomes. I also want to add human collaborators, deeper evidence grounding with citations, and feedback loops that can compare what the board recommended with what actually happened afterwards. The long-term idea is simple: not to replace the person making the decision, but to make sure they do not have to make every difficult decision alone.
Development disclosure
Junta Directiva AI existed as an earlier product concept before this hackathon. For the competition, I built and iterated the submitted Google Cloud implementation during the Submission Period: Gemini through Vertex AI, Google ADK orchestration, Cloud Run background execution, Firestore realtime state, Firebase Hosting, parallel agent activity, document context, Chairman follow-up and executive PDF outputs.
The public repository documents the architecture and includes setup instructions so the implementation can be reviewed and reproduced.
Log in or sign up for Devpost to join the conversation.