Chargebacks Suck >:(

Inspiration

The other day two of our group mates were talking with a startup founder and he was complaining about the tedious process of fighting chargebacks. Every startup or small business that sells physical items faces the challenge of chargeback resolution. These businesses receive hundreds of chargeback emails every day and typically founders spend hours daily resolving these chargebacks manually, collecting evidence from their inboxes, filling out forms, and writing emails. If businesses don't fight chargeback disputes, they could lose a majority of their revenue. We wanted to create a solution that could bring technology to the manual world of chargebacks and automate the process.

What it does

AI agents read the email inbox to scrape any potential chargeback notifications and potential evidence to dispute the chargebacks. The AI then puts the data in a dashboard and automatically assigns available evidence to each found chargeback case. Then the business owner can choose to dispute the case or not based off available evidence assigned. The AI will then format a chargeback rebuttal email and evidence to be sent to the customer's bank to refute the chargeback.

How I built it

The frontend is Next.js, database is hosted on Firebase, and backend is a Flask server. For the backend, we implemented a LLM chaining model called ReACT which allows us to run agents that can chain multiple tasks together. Our backend involves multiple asynchronous agents that process emails and computes a graph of evidence and dispute email nodes, extracting numerical features from each email using our model. The frontend then traverses this graph and allows the user to see which disputes have correlated evidence and can be resolved. The agents will then generate a dispute resolution email to fight the chargeback which can be sent to the merchant's bank.

Challenges I ran into

We had issues sourcing test data and had to manually create all the test chargeback data and invoice documents to simulate actual chargebacks based on real company templates. We also had issues with building the graph of evidence and dispute nodes due to inconsistency in data and disputes that share multiple similarities. We resolved this by implementing algorithms to post-process the output of the agents.