This is a product we built and run ourselves, not client work. We are showing it because it is the clearest proof of how Struxa builds: start with a real problem, ship something small that works, then expand it based on what users actually do.
The problem
Small shop owners in India run their businesses on memory and notebooks. Credit given to regular customers lives in a ledger or in the owner's head. Reminders are awkward to send. GST is confusing, and most accounting apps are tools the owner has to operate themselves, which is exactly what a busy shopkeeper does not have time for.
The approach: put the assistant where the owner already is
Instead of asking owners to download and learn another app, Munshio is a WhatsApp number. The owner messages it the way they would message a person, by text or voice note, in Hindi or English. There is nothing to install.
What we built
Five specialised AI agents sit behind that single chat:
Credit tracking and invoicing. Records who bought what on credit, straight from the chat.
Expense tracking and profit-and-loss, so the owner sees how the month is really going.
GST and tax guidance in plain language, with a hand-off to a real CA for filing.
Payment follow-ups. Drafts the reminder for the owner to send, and keeps track of promises.
Financial health checks and alerts when something in the numbers looks wrong.
How it works under the hood
The core idea is what we call a commitment graph. Every promise, deadline and obligation mentioned in a conversation (for example "I will pay on Friday") becomes a structured record. All five agents read from the same record, so the assistant remembers who promised what instead of treating each message in isolation.
- WhatsApp Cloud API for messaging, connected directly to Meta.
- FastAPI backend with an agent router that sends each message to the right agent.
- Claude to understand messy, informal messages and extract the details that matter.
- Indian-language speech models so voice notes work.
- Supabase (PostgreSQL) for the ledger and commitment records, hosted in India.
What early users taught us
We piloted with shop owners before building further. Two things changed the plan. They said they could manage basic credit tracking themselves, but wanted help with the harder part: following up on payments when customers go quiet. So payment follow-up moved up the list. And they were wary of fixed fees before seeing results, so we designed a free seven-day trial where the assistant works on pending payments first and the owner decides afterwards.
Design decisions that carry over to client work
- The owner stays in control. Customer-facing messages are drafted by the assistant, and the owner decides what gets sent.
- One workflow first. We shipped the credit workflow, then added agents, rather than building everything before anyone used it.
- Guardrails over cleverness. The assistant hands off to a human CA for filing and does not pretend to be one.
What this means if you want something built
The same approach works for client projects: WhatsApp-first, one workflow at a time, and a human always in control of what matters. If you want an assistant like this for your own business, see our AI automation packages, or start with a simpler WhatsApp ordering system.