Inspiration
Freelancers don't lose money because they can't do the work. They lose it in the gaps around the work: typing out invoices, double-checking the maths, and awkwardly chasing late payers. Most invoicing tools still make you fill in a form field by field. We wanted to see if a freelancer could just paste the message a client sent ("Bill Priya at Bluepeak, 12 hours of landing page design at $40/hr, plus a $60 flat fee for logo files, due in 10 days") and have a real PayPal invoice ready to approve.
What it does
InvoicePilot is an AI invoicing agent built on the PayPal Invoicing API.
- Read with AI: paste a plain-English message. Gemini extracts the client, business, email, line items, currency and due date.
- Review and approve: the draft appears in an editable form. Nothing is sent to PayPal until you approve.
- PayPal sends it: InvoicePilot creates and sends a real invoice through the PayPal Invoicing v2 API, with PayPal's hosted payer link.
- Live dashboard: paid, outstanding and overdue totals come from PayPal's own invoice data.
- AI payment reminders: one click drafts a friendly or firm reminder, editable before it goes out through PayPal's remind endpoint.
- Ask your invoices: questions like "Who owes me the most?" are answered from pre-computed totals, so the model cannot invent numbers.
How we built it
- Backend: Node.js and Express, with a small PayPal client (OAuth token caching, one automatic retry on auth failures) and a Gemini client that returns structured JSON.
- PayPal: Invoicing v2 in the sandbox: generate invoice number, create, send, list, remind and record payment.
- AI: Gemini 2.5 Flash with a response schema for extraction, reminder drafting and question answering.
- Frontend: vanilla JavaScript, no framework, so it loads instantly.
- Hosting: Vercel's free tier, deployed straight from GitHub. Total cost to run: zero.
Guardrails we added on purpose
AI that touches money needs limits, so the server does not trust the model:
- Totals are recomputed on the server from quantity and price; the model's arithmetic is never used.
- Emails and dates are validated, and missing information is flagged instead of guessed.
- Due dates default to net-14 when a message doesn't give one.
- Invoice totals above a cap are blocked.
- A human approves every invoice before it is sent.
- The Ask feature answers only from pre-computed facts about your invoices.
Challenges we ran into
- PayPal permissions: the first calls failed with a 403 until the sandbox app had the Invoicing feature enabled. We added a retry and a scopes diagnostic so the failure is obvious next time.
- Keeping the AI honest: the model could be asked about amounts it had no data for. Moving all arithmetic and totals to the server fixed that.
- Free hosting for a persistent server: we moved the Express app onto Vercel's serverless model, which meant exporting the app instead of calling listen directly.
Accomplishments that we're proud of
- A full loop from a pasted message to a sent PayPal invoice, then a reminder, then a paid invoice, all working on a live link.
- Real guardrails around AI and money, not just a demo prompt.
- Everything runs on free tiers.
What we learned
Structured output plus server-side validation makes an AI agent far more trustworthy than a clever prompt alone. And the Invoicing API suits agents well: every step (create, send, remind, record payment) is a clean, separate call.
What's next for InvoicePilot
- Live PayPal accounts and webhooks, so invoices flip to paid automatically.
- Recurring invoices and retainer clients.
- Voice input, and WhatsApp or email forwarding, as ways to submit the client message.
- Automatic reminder schedules, still with approval before sending.
Note for judges: the app runs against the PayPal sandbox, so no real money moves and sandbox invoices don't send real emails. Use Mark paid, or pay the invoice's payer link from a sandbox Personal account, to see an invoice flip to PAID.