--- title: Shopping Log Bot url: https://rone.dev/projects/shopping-log-bot description: A Telegram bot that reads receipt photos with Gemini and files them into a spreadsheet. language: en-US updated: 2026-09-27 translations: id: https://rone.dev/id/projects/shopping-log-bot --- # Shopping Log Bot A Telegram bot that reads receipt photos with Gemini and files them into a spreadsheet. [Projects](https://rone.dev/projects) / [AI](https://rone.dev/projects/category/ai) Live · Started 2025 · 4 GitHub stars · 0 Forks A twenty-node n8n workflow for logging personal spending. Photograph a receipt and a vision model extracts the store, total, items and date; type a sentence and an agent files it as an expense; send the report command and a second agent queries the sheet and answers in prose. Everything lands in Google Sheets — there is no application server and no database. ## Project facts | | | | -------- | --------------------------------------------------------------- | | Status | Live | | Category | AI | | Started | 2025 | | Role | Design and build | | Stack | n8n, Google Gemini, Google Sheets, Telegram | | Licence | MIT | | Links | [Source](https://github.com/ridwaanhall/n8n-pendataan-shopping) | ## Overview The friction in tracking spending is not the arithmetic, it is the typing. This removes it: the input is a photo taken at the till, and the interface is a chat app that is already open. A switch at the front of the workflow reads the incoming message and picks one of three paths — a photo, the report command, or free text. Each ends in the same spreadsheet, so there is one place the data lives regardless of how it arrived. ## Constraining the model instead of parsing it Receipt images are sent to the vision model with a prompt that specifies the exact four-line output format, gives a filled-in example, and defines a single error line for an unreadable receipt. The response is then read positionally. That is the trade being made deliberately: a rigid prompt and a dumb parser, rather than a loose prompt and a parser that has to interpret prose. When the model cannot read a receipt it says so in the one shape the workflow expects, which is a far better failure than a confident wrong total appearing in the spreadsheet. The two conversational paths work the other way round — they get an agent with the spreadsheet attached as a tool, because "how much did I spend on groceries last week" is a question, not a form. pip install rone-arena — the whole API surface behind four sub-clients. PyTorch LSTM forecasts for Indonesian gold, from one month to five years out.