Inspiration

Most budgeting apps tell you what you spent after the money is gone. By the time a bad month shows up in a chart, it's over. We wanted something that notices mid-month, explains the cause honestly, and helps you act while there's still time, without ever moving money on its own.

What it does

  • Detects spikes: projects each category's month (spent so far plus usual spending for the days left) and flags categories heading 30%+ and $25+ above a usual month.
  • Explains the cause: splits each increase exactly into more purchases, higher prices and subscription changes. AI writes the words; code supplies every number.
  • Plans with an AI agent: a tool-using agent combines savings options across the categories you marked as flexible, uses your record of plans kept and missed, tests combinations with a simulator, and proposes a plan with a reason for each choice. You can open every step it took.
  • Acts only with approval: when you approve, the server recomputes the amount, checks your PayPal balance, and moves the savings with a PayPal Payout that can never be paid twice. Then it tracks whether you keep the plan.
  • Alerts on your devices: Web Push notifications when a category starts spiking or a plan goes off track. No email, no app install.
  • Reads every card: PayPal Transaction Search, banks and cards through Plaid, or a CSV statement from any bank (the AI reads the layout, code reads every amount).

How we built it

  • PayPal (sandbox): Transaction Search to read spending, Balances to check before any transfer, Payouts to move approved savings (the approval id is the sender_batch_id, so retries can't pay twice), and Orders v2 for test checkouts.
  • AI: open-weight models through Groq: Qwen 3.8 for explanations, labels and statement layouts, and gpt-oss-120b for the planning agent (function calling). Gemini takes over automatically if the open model fails, and code-written text keeps everything working if both are down.
  • Server: Node.js 24 and Express 5, with SQLite (node:sqlite). Money is stored as integer cents.
  • Client: React 18 and Vite, an installable web app with a service worker for push.
  • Hosting: Render Blueprint with a persistent disk. Every visitor gets their own copy of the demo (a small SQLite file attached to a shared database), so judges never see each other's changes.
  • 133 automated tests and CI on every pull request.

Challenges we ran into

  • Keeping the AI honest with money. Models love to round, invent or re-calculate figures. We made code the only source of numbers: any figure in AI text that code didn't supply is rejected, the agent can only choose options code offered, and failed proposals go back to the agent with the reasons.
  • Open models and tool calling. Some models produced malformed tool calls or ran out of turns gathering facts one at a time. We give the agent a briefing up front, retry on rate limits, and restart cleanly on a backup model instead of switching mid-conversation.
  • A public demo that moves real sandbox money. Per-visitor workspaces, a daily payout cap and a per-visitor approval limit keep the shared sandbox account healthy through judging.

Accomplishments that we're proud of

  • 100% category and recurring-charge accuracy labeling the demo history (165 transactions) with Qwen 3.8.
  • The agent plans in about 5 seconds, and every number it shows is verified by code.
  • Approving a plan really moves money through PayPal, and the app follows up on whether you kept it.

What we learned

AI is most useful here as a planner and explainer, not a calculator. Splitting the work (AI chooses and explains, code computes and checks, the person decides) made the product both smarter and more trustworthy.

What's next for Budget Guardian

  • Real PayPal accounts (moving beyond the sandbox) and more currencies.
  • India: Account Aggregator for bank data, and UPI spending.
  • Shared budgets for couples and roommates.

Try it

No sign-up needed: https://budget-guardian.onrender.com. Each visitor gets their own copy of the demo; Settings → Reset the demo starts over. Plaid sandbox login: user_transactions_dynamic, any password.

Built With

Share this project:

Updates

Submission history