Inspiration

Receipts pile up in wallets, statements sit unread as PDFs, and budgeting apps want you to type everything by hand. Nobody does that. HackUTA's theme is Homer's Odyssey, and Penelope kept her household together for twenty years through patience and cleverness. Penny is named after her: a quiet, clever keeper of your money, making sure every penny finds its way home.

What it does

  • 📷 Scan a receipt with your phone camera. Gemini reads every item and sorts it into categories. A receipt with groceries and medicine is split correctly, with tax divided so totals match what you paid.
  • 📄 Import a whole bank statement PDF. Every transaction is read, card payments are skipped, and duplicates of receipts you already scanned are flagged.
  • 📊 A live dashboard: monthly and category spending, top merchants, a 3D voyage chart, and automatic recurring-charge detection (rent, subscriptions) with predicted next dates. Scan on your phone and the laptop updates within seconds.
  • 🦉 Ask Athena: "How much did I spend on dining in October?" Gemini writes the SQL, Penny runs it safely on your data, and Athena answers in plain English, even out loud.

How we built it

React + TypeScript (Vite, Recharts, Motion, three.js) on the front end, and Node + Express on the back end. TimescaleDB on Tiger Cloud stores spending in a hypertable. time_bucket powers the charts, and SQL window functions detect recurring charges. Gemini reads receipts and statements with strict JSON schemas, writes SQL for Athena, and falls back across models if one is rate-limited. ElevenLabs gives Athena her voice. Deployed on Vultr (Ubuntu, Dallas) behind Caddy with automatic HTTPS at allpenny.us.

Challenges we ran into

  • Making AI-written SQL safe: every query runs read-only, with a 5-second timeout and a row cap. We tested it against DELETE, DROP and hidden tricks.
  • The newest Gemini models sometimes never answered, and the free tier ran out mid-testing, so we built timeouts and automatic model fallback.
  • Money math: splitting tax across categories caused rounding errors until we did everything in cents.

Accomplishments that we're proud of

  • Snap a real receipt on a phone and watch it land on the dashboard seconds later.
  • 7 of 8 real test receipts read perfectly on the first try, including a 17-line receipt with discounts, matched to the cent.
  • An "ask your data" chat that's genuinely useful and can't damage the data.

What we learned

Structured output turns an LLM into a dependable data-extraction tool, TimescaleDB makes time-series questions short and readable, and AI features need guardrails: timeouts, fallbacks and read-only access.

What's next for Penny

Accounts and login, budgets with alerts, spending forecasts, TimescaleDB continuous aggregates for instant dashboards over years of data, and one-tap undo for imports.

Built With

Share this project:

Updates

Submission history