Inspiration
Splitting costs on a trip or in a shared house means a form for every coffee, a tangle of IOUs, and "I'll pay you later" money that never actually moves. We wanted logging an expense to be as easy as texting a friend, and paying back to take one tap.
What it does
- Talk: log expenses in plain English ("I paid $100 for concert tickets: Alex owes 60, I owe 40"), by voice, or by snapping a receipt. A note like "Alex had the burger" assigns receipt items to people, with tax and tip spread fairly.
- Split: exact to the cent, auto-categorized, and debts are simplified to the fewest possible payments ("3 payments instead of 6").
- Settle: every debt has a Settle via PayPal button. We create a PayPal order, the payer approves in PayPal Checkout, and we capture it. The debt disappears and every member's screen updates live.
- Understand: an AI copilot answers questions like "Who owes me the most?", an insights dashboard shows where the money went, and AI nudges write friendly payment reminders.
- A guided tour onboards new users in 60 seconds.
How we built it
- PayPal: OAuth2 client credentials plus Orders v2 (create with intent CAPTURE, then capture). Capture is idempotent, the order ID must match, and overpayment is rejected.
- AI: Groq (GPT-OSS 120B for text, Qwen 3.8 for receipt images). Every AI result is validated against a Pydantic schema, and all money math is done in code, never by the model. Models are auto-selected from the key's live model list, every call has a 25-second timeout, and an offline fallback keeps the app working if the AI is down.
- App: FastAPI + SQLAlchemy backend, React 19 + Vite + Tailwind frontend, Server-Sent Events for live sync. One Docker image, deployed on Render with a one-click blueprint. 39 automated tests and CI.
Challenges we ran into
Making LLM output safe enough to move money. Our receipt test showed the model reading every item correctly but getting the per-person math wrong, so we moved all arithmetic into code. AI models were also retired or overloaded mid-hackathon, so we added automatic model rollover, timeouts and honest fallbacks.
Accomplishments that we're proud of
One sentence becomes an exact split, then a PayPal capture, then a live update on every phone. Itemized receipts split per person from a casual note, correct to the cent.
What we learned
Treat the AI as a reader, not a calculator: let it understand messy human input, and let code guarantee the money is right.
What's next for SplitSync AI
User accounts and invite links, PayPal webhooks as a backup for capture, payouts straight to the friend who's owed, multi-currency with live exchange rates, and recurring expenses.
- Delegate: the Settle-Up Agent takes commands like "settle everything for me" or "pay back Sarah". AI works out the intent, the plan is computed from the ledger (the AI never invents amounts), and after you confirm it creates the PayPal orders and walks you through approving each one, plus writes reminders for anyone who owes you. It can only pay your own debts, and every amount is re-checked on the server.