555 Buku

Home: what's owed, overdue and collected at a glance
Paid through PayPal: the tab ticks itself off
The customer's own page, with a Pay now button
Collect: one tap sends each customer a PayPal invoice
Insights, with Nudge answering who to chase and drafting a reminder
Scan: AI reads the notebook page; the owner checks every line

Inspiration

Walk into almost any kedai runcit or kopitiam in Malaysia and you'll find a "buku 555": a small notebook where the owner writes down who took what on credit. "Ah Seng, 2 Bag of Chips RM10." It runs on trust, but pages get lost, names get smudged, and nobody likes asking a regular customer for money. Owners quietly lose hundreds of ringgit a month to tabs that are simply forgotten. We wanted to keep the notebook (it works, and owners trust it) and fix the part that doesn't: getting paid.

What it does

555 Buku turns a photo of a handwritten IOU page into money collected:

Scan: snap a page of the notebook. AI reads each line's name, item, amount and date, flags crossed-out lines, repayments and anything it isn't sure about, and the owner checks every line before anything is saved.
Collect: one tap sends each customer a PayPal invoice. Each customer also gets their own page showing what they owe, with a button to pay through PayPal. When PayPal says it's paid, the debt ticks itself off.
Remind: the owner sets the rules once (when to start, how often, when to stop, and never during Friday prayers). Reminders show on a calendar, can be dragged to another day, and go out as PayPal reminder emails or a one-tap WhatsApp message in the owner's chosen tone.
Nudge: an AI assistant that answers "who should I chase this week?" and drafts a polite reminder for a customer.
Ledger and Insights: an AG Grid ledger of who owes what, overdue first, and an AG Studio dashboard of totals and trends.
Works on a phone or a laptop, in nine languages (English, Malay, Chinese, Tamil and more).

How we built it

PayPal: the PayPal Invoicing API (v2, sandbox) creates and sends invoices, checks their status and sends reminder emails. PayPal webhooks (invoice paid, cancelled, refunded) update a Postgres database, so a payment shows up even while the app is closed. Owners can also sign in with Log in with PayPal.
Reading the notebook: Google Gemini (gemini-3.5-flash) reads the photo through the Vercel AI SDK and returns structured rows. Photos are kept in private Vercel Blob storage. The app then checks the model's work: dates are read day/month the Malaysian way, a missing year becomes the most recent such day, and anything unreadable is flagged for the owner instead of guessed.
The ledger and insights: AG Grid for the ledger, AG Studio for the dashboard, including a custom "who owes you" widget and an AI panel run by Nudge.
Reminders: a scheduled job runs every hour and sends due reminders at the owner's chosen time. Bryntum Scheduler and Calendar show the plan, with drag to move.
Nudge: Gemini through the AI SDK, with tools that read the ledger and draft messages; it never sends anything itself.
Stack: Next.js 16 and React 19 in TypeScript on Vercel, Neon Postgres, nearly 1,000 unit tests.
How we worked: a team of three sharing one GitHub repo, small pull requests reviewed before merging, and Claude Code as a pair programmer with a shared rulebook.

Challenges we ran into

Handwriting is messy, and so are dates. The model sometimes read "9/11" as September instead of 9 November, or guessed last year. We now take the date written at the start of a line ourselves and flag anything that isn't a real day.
Money needs to be exactly right. Retrying a failed batch must never send a customer two invoices, and a refund or cancellation has to undo a "paid" tick. We built per-invoice results, safe retry and webhook handling with tests for each case.
Keeping the owner in control. AI reads and drafts, but nothing is saved, sent or charged until the owner approves it.
Local details matter. Ringgit invoices only work between Malaysian PayPal accounts, and reminders should never land during Friday prayers.

Accomplishments that we're proud of

The whole journey works end to end in the PayPal sandbox: photo of a notebook, checked lines, PayPal invoices sent, the customer pays, and the debt ticks off by itself.
An app a corner-shop owner can use on their phone, in their own language, without changing how they already work.

What we learned

PayPal's invoicing does a lot of the heavy lifting (emails, a payer page, reminders), which let us focus on the owner's experience.
AI is great at reading and drafting, but it needs checks around it, especially for money and dates.
Small, reviewed pull requests kept three people (and an AI pair programmer) from stepping on each other.

What's next for 555 Buku

A monthly summary for owners ("RM2,340 collected, 3 regulars always late") that could also support a small-business loan application.
Reminders over WhatsApp and Telegram sent automatically, where customers already are.
A "talking till": the hawker says the order out loud and the customer pays by PayPal.
Testing with real shop owners in our neighbourhood.