catagory agent
monthy tax dedctaion agent
fast api

About the Project

Understanding personal finances is difficult—bank statements are messy, categories are unclear, and financial literacy varies widely. Many people simply don’t know where their money goes or which expenses may be tax-deductible.

Our goal was to create an AI-powered financial assistant that is simple, inclusive, and accessible for everyone. This led us to build FinSight AI, a multi-agent system that turns unstructured financial documents into clear insights using modern AI, RAG, and rule-based reasoning.

What Inspired Us

We observed a common problem: People struggle to interpret transactions, identify risk, and understand basic financial concepts.

We were inspired to build a tool that:

Helps users understand their spending

Flags unusual activity for safety

Highlights tax-deductible items

Improves financial literacy through simple explanations

Our motivation was to make financial clarity available to everyone — not just experts.

How We Built It

Ingestion Pipeline

We built a robust PDF-processing system:

Converts PDFs or images into text

Extracts structured fields:

transaction

{ date , merchant , category , amount } transaction={date, merchant, category, amount}

Stores everything in Parquet for fast analytics

Vector Store + RAG

We embedded:

Tax rules

Spending guidelines

Compliance patterns

Merchant information

Using Qdrant, the system performs semantic search to retrieve relevant financial rules. This allows the AI to answer queries grounded in real knowledge.

Multi-Agent Architecture

We created four specialized agents:

Classifier Agent – categorizes each transaction

Tax Agent – identifies deductible expenses

Compliance Agent – detects suspicious or risky patterns

Summary Agent – computes monthly totals and spending analytics

These agents are coordinated by a Dispatcher Agent, ensuring each task is routed efficiently.

Streamlit Frontend

We built an accessible UI that:

Lets users upload PDFs

Displays clean transaction tables

Provides insights and summaries

Allows natural-language questions

Supports inclusive, user-friendly interaction

Challenges We Faced

Cleaning noisy PDF data into a reliable format

Designing agent logic that stays consistent

Integrating RAG to improve accuracy, not hallucinations

Keeping latency low inside Streamlit

Ensuring the system works for many document formats

What We Learned

Multi-agent systems are powerful when each agent has a clearly defined purpose

RAG significantly improves financial reasoning when grounded in solid rules

Inclusive design and simple UI/UX matter more than expected

Building production-style pipelines in 24 hours requires prioritizing clarity and modularity

Financial literacy tools must be accessible to all backgrounds

Final Thoughts

Our mission was to create a tool that not only processes numbers but empowers people. FinSight AI helps users understand spending, identify risks, and make smarter decisions — all through an accessible, AI-driven experience.

We believe this is a step toward making financial clarity and education available to everyone, regardless of background.