I recently participated in a hackathon focused on building AI agents. The main rule was clear. AI could not be the product by itself. It had to support a real system.
Our team had five members. We worked for just over 24 hours. The pressure was real. So was the ambition.
The Problem We Chose
Many people have business ideas. Very few know if those ideas make financial sense before investing money.
We saw a gap. People needed a way to test ideas early, using numbers and logic, not hope and vibes.
Our goal was to build a tool that helps someone answer a simple question. Is this business idea viable or not?
The Core Idea
We built a business viability checker called Microbiz.
The user enters structured inputs. These include starting capital, pricing, expected monthly sales, operating costs, and time horizon.
At the core of the system is a calculation engine. It computes revenue, total costs, profit margins, and sustainability indicators over time.
This part was fully deterministic. Just math and logic to get consistent results every time.
We wanted users to trust the numbers first.
System Design and Team Roles
We used Django for the back-end. I worked mainly on this part.
My role involved handling inputs, validating data, running financial calculations, and structuring outputs. I also prepared the data so it could be consumed cleanly by the AI component.
Two teammates worked on the front-end. Their focus was speed and clarity. They designed an interface that made complex inputs easier to manage under time pressure.
Another teammate focused on prompt engineering. Their job was to control how the AI reasoned, what context it received, and how it communicated results.
Where the AI Fit In
We integrated Groq AI using an API key.
The AI did not redo the calculations. It received the raw user inputs plus all internal computed values.
From there, it performed higher level reasoning. It considered market conditions, pricing realism, demand signals, and feasibility based on what is currently visible online.
The AI produced one output; it gave a recommendation on whether the business idea was likely viable or risky under the given assumptions. It stated reasons based on the input given
The AI was meant to support decisions, not motivate or sell dreams.
The Reality Check
We underestimated the workload.
Writing formulas was not the hardest part. Defining realistic assumptions was.
We had to research small business behavior, cost structures, margins, and failure points. That research slowed development since none of us have any financial or business backgrounds, but skipping it would have made the system useless.
Time worked against us. Features piled up faster than we could finish them. We created and broke code again and again until we got to a middle ground that wasn't perfect but works well enough.
Despite working through the night, we did not complete everything we planned.
Pitch Day
Before demos, all teams were called to a common area and given three minutes for an elevator pitch with slides and demos. When our turn came, we presented with nothing but words.
It still worked in our favor since we understood the product deeply. We explained the problem clearly and answered the judges' questions without much thinking. We explained why pure AI was not enough and how logic and AI reasoning worked together.
We did not make the podium. However, the judges told us the product was useful and encouraged us to keep building it. That feedback mattered more than placement.
Lessons Learned
First, scope matters. AI agents amplify systems. They do not replace solid foundations.
If your assumptions are wrong, AI makes the output confidently wrong. It's just how systems work; GIGO (Garbage In Garbage Out)
Second, research is not optional. Especially when dealing with money and real people.
Third, clear ownership helps under pressure. Back-end, front-end, and AI roles stayed mostly separate. That prevented chaos.
What Comes Next
This project is not finished and it most definitely won't be shelved.
Our next step is reducing the amount of input required. The system should do more inference with less user effort so that even those who are green in business but would still like to start one can do it but with more knowledge.
Long term, the goal is idea discovery. The system should recommend business ideas directly, starting with specific regions in Kenya. And as the scope expands maybe even worldwide. Maybe one day you may even interact with it.
The hackathon may have ended, but the product will live on.
Top comments (0)