CashFlux

Home Page
Dark Mode
Receipt Scanning
Project Proposal

Why CashFlux

Every month, the CEO of a 50-person company drowns in hundreds of transactions with no visibility, no context, and no way to catch problems before they become expensive. We wanted to build the financial watchdog that small businesses can't afford to hire.

The Product

CashFlux is an AI-powered expense intelligence platform. You ask it anything about your company's spending in plain English and it answers instantly with charts, summaries, and insights.

Behind the scenes it flags policy violations, tracks employee spending credit scores, detects geofencing anomalies, and runs a fraud detection engine that catches suspicious patterns across cards, including patterns invisible to the human eye.

How It Works

We built a multi-turn AI agent using Google Gemini Gemini 2.5 Flash with function calling, connected to MongoDB MongoDB Atlas as the data backbone.

The agent dynamically decides which database queries to run, retrieves real transaction data, performs analysis, and generates charts automatically. The fraud detection engine is a rule-based scoring system built around 10 independent signals, including cross-card device reuse, merchant burst detection, velocity attacks, geographic anomalies, unusual spending patterns, and other indicators commonly associated with fraudulent activity. By combining these signals, the system produces an interpretable risk assessment that helps reviewers understand why a transaction was flagged.

The backend is built with FastAPI FastAPI and exposes the services used by both the AI agent and the review platform. The reviewer interface is a keyboard-driven single-page application designed for fast investigation workflows, allowing analysts to review alerts, inspect evidence, and make decisions efficiently.

To create a more natural and interactive experience, we also integrated ElevenLabs ElevenLabs for voice synthesis. This enables the agent to communicate findings through realistic speech, making it possible to interact with the system using voice while receiving clear, human-like explanations of detected fraud patterns and risk factors.

Challenges

Getting the Gemini function-calling loop to work reliably with multi-turn conversation history was tricky.

We also had to discover the 4 fraud patterns in the dataset through data analysis with no labels, finding the merchant burst pattern required cross-card aggregation.

Wins

The agent genuinely answers follow-up questions without re-explaining context.

We are also proud of the fraud reviewer UI. It works entirely with keyboard shortcuts, a non-technical person can triage flagged transactions in seconds.

Lessons

Building an explainable system is harder than building an accurate one, every flag needs a reason a human can act on.

Next Steps For CashFlux

Receipt scanning via Gemini Vision, one-click expense approvals with ElevenLabs voice narration, real-time transaction streaming, and training a model on reviewer decisions to continuously improve detection accuracy.