A lightweight offline tool uses Gemma 4 to interpret Hindi, Marathi, Hinglish and English orders while keeping pricing, customer data and payment details on the shopkeeper’s machine.
Parchi is a lightweight billing tool aimed at the everyday chaos of neighbourhood kirana shops, where WhatsApp messages arrive in a mix of Hindi, Marathi, Hinglish and English and must be turned quickly into accurate bills. The app lets a shopkeeper paste a message, or open an exported chat file and tap a line, then uses Gemma 4 on the shop’s own laptop to identify the products, quantities and size words before matching them to the store’s price list. Within seconds, it produces a checked order, a printable packing slip, a WhatsApp reply and a UPI QR code for the exact amount, all without sending customer data to a remote server.
That local-first design is at the centre of the project. According to the developer’s description, the model is only responsible for reading the message; pricing and totals are handled by ordinary code on the machine itself. The system uses structured output so the response is predictable, then applies a matching layer that tries exact aliases first and falls back on fuzzy matching when needed. It also handles common quantity expressions such as “aadha”, “dedh” and “paav kilo”, converting them into standard units before the bill is calculated. If an item does not fit neatly into a pack size, the line is highlighted for the shopkeeper to decide rather than being guessed automatically.
The project also sits within a broader push towards software built for small Indian retailers. Companies such as Tillvio, WapiGrow, KiranaFlow and BillDost already market billing and shop-management tools for kirana stores, often with WhatsApp-based receipts and offline support. Parchi differs by focusing narrowly on the first, repetitive step in the workflow: interpreting the incoming order itself. Its creator argues that this matters because kirana shops work on thin margins, handle private household buying patterns and cannot afford per-message API fees for dozens of small orders a day. A locally run model also reduces reliance on patchy connectivity, which remains a practical constraint for many small businesses.
The developer says the current build runs on FastAPI and SQLite, works offline after the model is downloaded and avoids external fonts or content delivery networks. A local evaluation covering 20 orders and 34 item lines reportedly produced correct product and quantity matches across Hinglish, Marathi, Hindi and English, with the “thank you” non-order example correctly ignored. The same write-up acknowledges the limits: it is not a GST invoice system, it handles one message at a time and it is still text-only. Even so, the project presents a clear use case for open-weight models in small retail, where speed, privacy and low operating cost matter as much as accuracy.
Disclaimer: This article is intended to inform and educate, not to recommend or endorse any financial product, investment or strategy. Please consider your own financial circumstances and seek professional advice where appropriate before making financial decisions.
