Earlier this month (8th March), we participated in an Agentic hackathon, a competition focused on building AI systems where LLMs act autonomously to make decisions. The event occurred at The Beacon and challenged participants to create innovative multi-agent AI applications within a single day (6 hours of development).
Together with Tan Dalkiran, a talented developer (and big friend) studying in Eindhoven, we built FAInance — a web-based AI platform that automates financial document analysis using an intelligent multi-agent system. Our project won for its technical complexity and multi-agent coordination; not only did we implement an agent to analyze every CSV company file, but also one to search online for similar company data and another AI to judge the AI processes, and the last one to create a final result to show in a web client dashboard.
The Problem We Tackled
The idea for FAInance stemmed from the bureaucratic complexity of analyzing company balance sheets and financial documents, especially in Italy. Businesses often struggle to extract meaningful insights from financial reports due to the overwhelming amount of data and industry-specific benchmarks.
We aimed to automate and simplify this process, creating an AI-driven system that not only analyzes financial documents but also compares company performance against industry trends to provide actionable recommendations.
How FAInance Works
Users upload a CSV or Excel file, containing financial data such as revenue, expenses, and assets. The backend then activates five AI agents, each specializing in a different task:
🔹 Data Understanding Agent — Identifies columns, their purpose, and the meaning of the data.
🔹 Statistical Analysis Agent — Performs key calculations to assess company performance.
🔹 Market Research Agent — Analyzes similar companies, industry trends, and competitive advantages.
🔹 Report Generation Agent — Compiles insights into an HTML report and generates a JSON file for interactive visualizations. �
🔹 Coordinator Agent — Manages the workflow and ensures smooth interaction between agents. Acts also as a judge in case the output of the previous agents isn’t good enough.
One key feature of FAInance is that users can see the reasoning process of each agent, making the decision-making transparent. (All this process can be actually seen by the user from the console outputs)
The Tech Behind FAInance
Our stack was designed to maximize efficiency and scalability:
✅ Flask — Backend API and server-side processing.
✅ Vue.js — Interactive frontend to display results.
✅ CrewAI — Multi-agent system management.
✅ OpenAI & Gemini — Used together to optimize performance;
OpenAI handled high-neural-capacity tasks, while Gemini processed large files with its 1M token context window.
The biggest technical challenge was making the five agents interact cohesively without bottlenecks or redundant processing. We solved this by structuring their communication flow and optimizing API calls.
CrewAI: The Real Star of the Application
At the core of FAInance is CrewAI, a framework that enables the creation and coordination of autonomous AI agents. We designed our system using CrewAI’s declarative approach, defining specialized agents and assigning them structured tasks.
- Data Cleaning Agent: This agent is the first step in the workflow. Its main responsibility is to preprocess the uploaded CSV or Excel files — cleaning and formatting the data. It extracts an initial snapshot (for example, the first 10 rows) to provide a clear overview of the dataset’s structure. This step is crucial for identifying key columns and ensuring the subsequent analyses work on well-organized data.
- Statistical Agent: Once the data is cleaned, the Statistical Agent takes over. It performs comprehensive statistical calculations on the financial data. This includes computing key financial metrics such as total revenue, net profit, and other critical performance indicators. Its analysis transforms raw numbers into meaningful insights that help evaluate the company’s financial health.
- Internet Searcher Agent: This agent conducts online market research. It uses integrated search tools to gather information about similar companies and industry trends. By comparing the current company’s data with external benchmarks, the Internet Searcher provides valuable context and recommendations. Its role is essential for understanding the competitive landscape and identifying areas for improvement based on others similar companies.
- Reporting Analyst Agent: The final piece of the puzzle is the Reporting Analyst Agent, which compiles all the information gathered into a cohesive report. It synthesizes the data from the previous agents and formats it into an easily digestible HTML report. This agent ensures that every insight — from statistical figures to market research — flows together seamlessly, offering users a comprehensive view of their financial analysis.
- Judge Agent. (to implement)
This agent will act as an overseer and mediator, reviewing outputs from the specialized agents to ensure all critical analyses are complete and coherent. If an agent’s output appears incomplete or requires additional context, the Judge Agent will prompt for further details or request re-analysis, thus ensuring the final report meets the highest standards of accuracy and depth.
internet_searcher:
role: >
Market Research Specialist
goal: >
Based on the current company information, search for similar companies that are known to be succesfull
to give the current company financial and business advices, while keeping in mind the current company data.
backstory: >
You're a meticulous analyst with a keen eye for detail. You are known by your market search abilities, that
allow you to find and analyze similar companies. You are also gifted with the ability to see what your current
company and the researched companies are doing differently such that your company is failing to do.
Why We Won
The multi-agent complexity of fAInance set it apart from other projects. While many teams built single-agent applications, we demonstrated how AI agents can collaborate to solve a real-world business problem. Judges were particularly impressed by:
✔️ The intelligent coordination between five specialized agents.
✔️ The industry comparison feature, which helps businesses benchmark performance.
✔️ The transparent “thinking process”, making AI recommendations more trustworthy.
What’s Next?
One of FAInance’s most promising features is its ability to compare company performance against real-world financial sheets to provide cash flow improvement suggestions. This could be further developed to offer personalized financial strategies for businesses.
This hackathon was a great experience, and we’re excited about the potential of FAInance beyond the competition!
🚀 Would you use AI to analyze financial reports? Let us know your thoughts!
We are both students, ready to embrace new knowledge and experiences, feel free to reach out to us anytime!
Luis: https://www.linkedin.com/in/luis-beqja/
Tan: https://www.linkedin.com/in/tan-dalkiran/