TrackIt

Inspiration

I moved to New York City a month ago to live with a friend. I budgeted my food and transportation expenses at $600/month. I initially thought of using an Excel sheet to monitor my spending, but I quickly realized that I could vibe code an app that takes my receipts, which would be much faster than maintaining an Excel sheet. I naturally thought of integrating all my bank accounts into my app to show aggregate spending across all my credit cards, checking, and savings accounts. While living on a budget, I also wanted to reduce my spending and have been thinking of making agents that accomplish just that. Today, I built two such agents and plan to build two more.

What it does

Connects to all your savings, checking, and credit card accounts and aggregates them into a list and a pie chart, and you can select a date range to reveal total spending in that range and the categories (e.g., transportation, shopping, food) of those items. The first agent I have built helps find all the subscriptions you are subscribed to and can help you unsubscribe from the ones you don't use. The second agent helps you recognize which card, of all the credit cards you have, you spend at what stores and what items to help you get more cash back. The third (future) agent would help with searching and providing alternative options that are cheaper for you to purchase goods and services, and the fourth (future) agent would help with identifying all the fees and interest you're paying, like overdraft, ATM, and late fees and credit card interest, and help you get them waived or refunded.

How I built it

I designed the high level architecture myself and chose the llm models, languages, frameworks, deployment and database tools and any other api required, considering the tradeoff of each of these. Then asked claude to write a detailed implementation plan, asking for any correction it might need. I auto automated unit and regression testing. I have been working on this project for about 2 weeks now. For each new feature, I describe in detail what I want I want and Claude makes a plan for implementation which I review and implement the feature.

Challenges I ran into

The hardest part was how messy real bank data is. The transaction names the banks send are cryptic, and it was surprisingly tricky to tell apart real spending from money movement like Zelle transfers, credit card payments, and paychecks. At one point the app happily said all my accounts were synced, but my Bank of America transactions were missing, and it turned out I was silently filtering most of them out, so I had to rethink how the app decides what actually counts as spending. Categorizing items consistently across receipts, manual entries, and bank transactions with one set of categories was also harder than I expected, and making sure one person's financial data could never leak into someone else's took real care.

Accomplishments that I'm proud of

I'm proud that this is a real app connected to my actual bank accounts and running in production, not just a demo. In about two weeks I built the core spending tracker plus two working agents, one that finds all your subscriptions and helps you cancel the ones you don't use, and one that tells you which card to use where to earn more cash back. I also set up automated testing and a deployment pipeline so every new feature ships without quietly breaking something that already worked, which let me keep moving fast with confidence.

What I learned

I learned just how messy real financial data is, and that the right move is to let the LLM handle the fuzzy parts like reading receipts and categorizing, while keeping the actual math simple, deterministic, and easy to test. I also learned that just showing people their spending isn't worth much anymore, and the real value is in agents that act for you, like canceling a subscription or getting a fee refunded. And working with Claude, I found a rhythm where I describe what I want in detail, review the plan it writes back, and then build it, which let me ship far more in two weeks than I could have on my own.

What's next for TrackIt

Build out the next two remaining agents. Use the app by myself for the next two weeks, improving UI and fixing any bugs in between. I will then launch the app on TestFlight to my friends so they can use it and give me feedback. Depending on how that goes, I would add or remove features and finally submit it to the App Store.

Built With

• amazon-web-services
• fastapi
• gemini
• plaid
• python
• react
• shadcn/ui
• sqlmodel
• supabase
• tavily
• typescript