Inspiration

I kept coming back to the same thing: every small business owner I know in Lagos, shop owners, salon operators, market traders, does their books the same way. In their head, or on scraps of paper, at midnight, after the shop closes. They can't afford a bookkeeper. They lose track of who owes them money. They run out of stock without warning. And now, with Nigeria's new 2026 presumptive tax regime just taking effect, most of them don't even know if it applies to them. I wanted to build something that didn't ask them to learn a new app or change how they work, something that just showed up inside the one app they already have open all day: WhatsApp.

What it does

BizAgent is an AI agent that runs the operational side of a small business through WhatsApp. An owner texts, sends a voice note, or snaps a photo of a receipt, and the agent logs sales and expenses, tracks customer credit, manages inventory with low-stock warnings, prepares supplier payments for the owner to confirm before anything is sent, and answers real questions about the business by reasoning over its actual data, not a script. It also gives a plain-English estimate of where the business stands under Nigeria's new tax rules. Every business gets a web dashboard too, with the same AI agent built directly into it, so the data lives in one place no matter how you talk to it

How I built it

The whole system runs on Node.js and Express, with Gemini as the decision-making core. Every incoming message, whether from WhatsApp or the website, goes through one shared function that classifies intent and decides what action to take: log a sale, flag low stock, draft a payment, or just answer a question. Supabase holds all the data, and that same database backs the WhatsApp bot, the website's live chat demo, and the dashboard, so nothing is ever out of sync. Twilio handles WhatsApp delivery, and Paystack is wired in for real supplier payments, though the agent never sends money without the owner explicitly confirming first. The whole thing is deployed on Render behind a proper domain, with the dashboard PIN-protected and the tax and inventory logic all running on the owner's real, live transaction history.

Challenges I ran into

The biggest lesson was that reliability problems rarely look like what they actually are. A missing Supabase schema cache update looked like a broken feature. A stray git repository at the wrong folder level nearly overwrote an unrelated project of mine before I caught it. Google deprecated the entire Gemini SDK I'd built on partway through, which meant migrating the whole AI layer mid-build. And Twilio's free WhatsApp sandbox requires every single new number to send a manual join code before it works at all, which is fine for a developer testing alone but is real friction for an actual first-time customer. Instead of accepting that friction, I rebuilt the onboarding experience so the same real agent lives directly on the website too, with no WhatsApp setup required to try it.

Accomplishments that I'm proud of

The thing I'm most proud of is that this doesn't feel like a demo that only works in one narrow scripted path. A real business owner can talk to it however they naturally would, in English or Pidgin, by text or by voice, and it holds up. The dashboard, the WhatsApp bot, and the website chat all genuinely share one brain and one source of truth, not three separate things stitched together for a pitch.

What I learned

I learned that building an agent people will actually trust with their business records means being conservative exactly where it matters most, which is why the agent will draft a payment and explain it clearly, but it will never move money without a human saying yes first. I also learned how much real financial literacy support is missing for small businesses here, most owners have never had anything explain their own numbers back to them in plain language before.

What's next for Untitled

want to get BizAgent in front of real business owners beyond this submission, move off the WhatsApp sandbox onto a full production number, and build out staff logins so an owner's shop attendant can log sales too when the owner isn't the one at the counter.

Built With

Share this project:

Updates