Landing page: paste a WhatsApp order chat and the AI turns it into an order you can check.
Sign in with your email and password.
Turn on 2FA with any authenticator app by scanning a QR code.
Two-factor sign-in: enter the 6-digit code from your authenticator app.
Dashboard: today's orders and sales, cash to collect, busiest area, and AI tips based on your own numbers.
New order: paste a chat, try a Roman Urdu or English sample, or upload a screenshot or voice note, then press Read with AI.
Orders: every saved order with the customer, area, total and status.
Dispatch: pending orders grouped by area, with the exact cash the rider should collect.
Khata: see who owes you money at a glance, with filters for dues, paid and credit.
Customer khata: balance due, and a WhatsApp payment reminder written by AI in Roman Urdu or English.
Payment history for a customer.
Record a payment by Cash, JazzCash, Easypaisa or bank transfer.

Inspiration

Walk through any Pakistani WhatsApp or Instagram shop and you will see the same thing. Orders come in as messy Roman Urdu messages like "bhai 3 kurtay chahiye medium, 1800 ka ek, 1000 advance bhej diya", as screenshots and as voice notes. The seller copies each one into a notebook by hand. Advances get forgotten, nobody is sure who still owes money (the khata), and the rider leaves without knowing how much cash to collect. We wanted to take that daily headache away.

What it does

TradeFlow AI is a simple back office for these sellers.

Turn a chat into an order. Paste the customer's message, upload a WhatsApp screenshot or a voice note. The AI fills in the name, phone, address, items, prices and advance. Anything it is unsure about is shown in yellow, and nothing is saved until the seller checks it and presses Confirm.
Keep orders, receipts and deliveries in one place. Every order is saved to the seller's own account. Receipts can be printed, and the dispatch list groups orders by area and shows exactly how much cash to collect.
Track the khata. See what each customer owes, record payments by Cash, JazzCash, Easypaisa or bank, and recognise repeat customers even when they write their number differently.
Write WhatsApp messages with AI. One tap writes an order confirmation or a polite payment reminder in Roman Urdu, Urdu or English, ready to send.
Get tips from your own numbers. The dashboard turns the shop's best-selling items, busiest areas and biggest dues into three practical tips.
Stay secure. Sign up, sign in and protect your account with two-factor authentication (2FA). Each seller only ever sees their own data.

How AI is used

AI sits at the heart of the main journey, not on the side.

Google Gemini reads the chat, the screenshot or the voice note and returns the order in a fixed format. We taught it how people really write in Roman Urdu: number words like do, teen, darjan, dhai hazar, 1.5k, sizes and colours, delivery charges, and "advance nahi diya".
We never trust the AI blindly. Every answer is checked before use and retried if it is wrong. Each field gets a confidence score, and unsure fields are highlighted so the seller can fix them.
If Gemini is busy, the app automatically tries a backup model, so the seller is not left waiting.
The WhatsApp messages and dashboard tips are written only from the seller's saved numbers, never made up. All money is calculated by the app itself, never by the AI.

How we built it

We started by writing the plan for the AI agent: a product spec (SPEC.md), rules for the agent (AGENTS.md) and reusable skills (SKILLS.md and .agents/skills). Then we built the app step by step with the Devmarketplace builder agent: the database and its security rules, sign-up and 2FA, the AI order reader, orders, khata, dispatch, receipts, screenshots and voice notes, AI messages, AI tips and the final design. After every step we tested the live app with real Roman Urdu chats and fixed what did not work.

Built with: Next.js and TypeScript, Tailwind CSS, Supabase (Postgres database, sign-in with 2FA, row-level security), Google Gemini, zod for checking AI answers, and Vercel for hosting.

Challenges we ran into

Roman Urdu has no fixed spelling (kurta / kurtay, dhai hazar / 2500 / 2.5k), so the AI needed lots of examples and strict checks.
The rider must never be asked to collect more than the customer really owes, even when payments were recorded separately, so we kept all money rules in one tested place.
Gemini was sometimes overloaded, so we added time limits and backup models.

Accomplishments that we're proud of

A complete journey that works on the live site: sign up, set up 2FA, turn a chat or screenshot into an order, save it, and find it again after signing back in or restarting the app.
Four reusable code gigs that came out of this build:

What we learned

Small sellers do not need magic, they need AI they can check. Showing how sure the AI is, keeping the seller in control and letting the app do the maths made the AI something people can actually trust with their money.

What's next for TradeFlow AI

Taking orders straight from WhatsApp Business, Urdu-script receipts, low-stock alerts and a simple app for riders.

Try it: https://tradeflow-ai-phi.vercel.app (create an account, set up 2FA, then New order, pick a Roman Urdu sample and press Read with AI) Code: https://github.com/amnaatgit/tradeflow-ai