Inspiration
I was researching financial disputes and discovered the Consumer Financial Protection Bureau sent over 1.3 million complaints to companies in 2023. The manual review process for credit card disputes seemed like an interesting problem to tackle with AI. While I am sure sophisticated systems exist in the industry, I wanted to build my own take on this challenge - creating an AI that balances fraud detection with customer sentiment analysis and business impact calculations. It was a great way to explore AI applications in finance and demonstrate my coding abilities in a hackathon setting. Sources: https://www.usa.gov/agencies/consumer-financial-protection-bureau https://www.consumerfinancemonitor.com/2024/04/05/cfpb-publishes-consumer-response-annual-report/
What it does
This is an AI-powered assistant that helps review financial disputes in real-time. It takes a complaint or case description and runs it through three AI engines: -Fraud Detection Engine -Sentiment Analyzer -Predictive Engine(business reports)
How I built it
Built with React frontend and Python Flask backend integrating Perplexity Sonar API as the core intelligence engine. Used Sonar Pro model with custom prompts designed specifically for financial dispute analysis , teaching it to think about fraud patterns, customer emotions, and business costs simultaneously. Frontend: React + Vite Backend: Python Flask API AI Engines: custom-built logic Caching: Redis Environment Management: Used .env files and secure variable loading Version Control: Git + GitHub
Challenges I ran into
The hardest part was teaching Perplexity to think like a banker, not just a fraud detector. I had to write really specific prompts to make it understand. Getting the API calls to work smoothly was also little tough. Perplexity takes 8+ seconds to respond, so I had to make sure users knew it was actually working and not just frozen.
Accomplishments that I am proud of
I'm really proud that I got this whole thing working! Perplexity actually gives smart business advice, not just "fraud or no fraud." Like when it sees a suspicious charge but says "just give the money back because the angry customer will cause more problems." The good part is everything is connecting properly and when the perplexity API is called it is giving back the result.
What I learned
This was actually my first time building anything with LLMs, so I learned lot of things! Getting React to talk to Flask to talk to Perplexity and making it all work together smoothly was honestly pretty challenging but really rewarding.
What's next for Perplexity FincaseAI
I want to keep experimenting with the prompts to make Perplexity even better at this kind of analysis. Maybe try connecting it to real banking data or add features that help customer service people actually use this in their daily work. Since this was my first LLM project, I'm excited to try building other stuff with AI. The whole idea of making AI that understands context and emotions could work for so many different industries and problems.
Log in or sign up for Devpost to join the conversation.