Inspiration
Think of the home baker down the street, the tutor, the babysitter, the woman who braids hair on weekends. Their business runs on trust: "Take the cake, pay me Friday." Many of them are immigrants running a business in their second language.
Three things go wrong for them:
Nobody writes it down. Invoicing software is built for accountants, not for someone with flour on their hands. Asking for money is awkward, especially from neighbours, and harder still in a second language. So people don't ask, and the money quietly never arrives. The bills arrive anyway. Rent is due Friday whether the customer paid for the cake or not. We wanted to build the thing that does the awkward part for them, on top of the payment system their customers already trust.
What it does
Tab is a phone-first web app. The owner talks; Tab does the paperwork and the chasing through PayPal.
Say it. Tap the big button and say "Priya took three dozen cupcakes, she'll pay Friday." Any language works: Spanish, Italian, Hindi, Tamil and more. Check one sentence. Tab shows "Invoice Priya $90 for 3 dozen cupcakes, due Friday", plus a preview of exactly what Priya will receive, written in her language in the owner's own voice, with an English translation underneath. Send with PayPal. A real PayPal invoice with a Pay button goes out. There's a 5-second undo, and a one-tap option to also send it on WhatsApp, where these customers actually talk. "Today" card. Tab projects cash over the next 14 days against upcoming bills: "You're $450 short for Rent on Friday." It then proposes a plan: a gentle nudge for one customer, a firmer reminder for someone 9 days late, an offer to split the bill for someone who seems stretched, and leave Priya alone, she was reminded today. Nothing is sent until the owner taps Send. Follow-ups through PayPal. Reminders go through PayPal's reminder API. "Offer to split" turns on partial payments on the invoice, so the customer can pay half now with the same link. Got paid in cash? One tap records the cash payment on the PayPal invoice, so the books stay right. Money lands. When the customer pays on PayPal, the row turns green, cash goes up and the forecast updates live. Books. The full ledger lives in an AG Grid table with filters, search and CSV export. The AI never moves money or sends a message on its own. It understands, writes and plans; the owner approves.
How we built it
PayPal, the core of the product: Invoicing API v2: create, send, remind, cancel, update (to turn on partial payments) and record payment (cash). Webhooks (INVOICING.INVOICE.*, signature-verified) with polling as a fallback. Transaction Search for recent income. Every POST/PUT carries a deterministic PayPal-Request-Id, so a retry can never create a duplicate invoice or a second reminder. AI with structured outputs. Every AI step returns JSON that matches a Zod schema and is validated again on our side. There are three AI jobs: Voice note → invoice: matches known customers, resolves "Friday" or "lunes" to a real date, uses the customer's usual price if none was said, and asks instead of guessing. Messages: invoice notes, nudges, firmer reminders and split offers in the customer's language, copying the owner's tone from a sample of how they text. Collections planner: reads cash, bills, open invoices and each customer's payment and reminder history, and returns a prioritised plan with a plain-English reason for each action. Models: Google Gemini (gemini-3.6-flash, with an automatic switch to gemini-3.5-flash-lite when busy) by default. Claude (claude-opus-5-5) is supported as a drop-in alternative. If the AI is unavailable, Tab falls back to simple rules so the app never breaks mid-demo. Voice: the browser's built-in speech recognition. App: Node.js + Express, a single-page phone-first UI, Server-Sent Events for live updates, and an installable web app manifest. AG Grid Community for the Books ledger. Render for hosting, deployed from a render.yaml blueprint. We used Claude Code as a coding assistant while building.
Challenges we ran into
Free-tier AI capacity. Our first Gemini model returned "high demand" errors and once hung a request. We added timeouts, an automatic backup model and a rules fallback. Dates are surprisingly hard for an AI. The planner once called Friday "Thursday." We now give it the weekday names directly instead of letting it work them out. PayPal's status for cash payments is MARKED_AS_PAID, not PAID. We normalised it so the owner just sees "Paid." Making it not feel like a dashboard. Our first version put everything on one screen. We rebuilt it phone-first around one number, one card and one big button.
Accomplishments that we're proud of
A Spanish voice note becomes a real PayPal invoice with a warm Spanish message in about 10 seconds. The planner makes human judgement calls, like don't nag someone you reminded today, and explains them in one sentence. Real PayPal depth: invoices, reminders, partial payments, recorded cash payments, webhooks and idempotency, not just a checkout button.
What we learned
The best AI features are invisible. Owners don't care that it's AI; they care that the awkward message got sent, kindly, in the right language. PayPal Invoicing already does much of what small sellers need. Their real problem is that nobody has time to use it.
What's next for TAB
WhatsApp Business API, so reminders go out automatically on the channel customers actually read. Accounts and a real database, so many owners can use it. Learning each customer's payment habits over time to time nudges better. PayPal production onboarding for real sellers.
Built With
- ag-grid
- claude
- express.js
- google-gemini
- node.js
- paypal
- paypal-invoicing-api
- paypal-transaction-search
- paypal-webhooks
- render
- web-speech-api
- zod