LocalReceipt AI：本地收據開支分類工具
LocalReceipt AI
LocalReceipt AI 將貼上的收據文字轉換成結構化開支資料，協助自由工作者及獨立專業人士整理帳目。這個本地原型以 Qwen2.5 3B、Ollama、Python 和 Flask 建置，在作者的 Mac 上運行，並非公開部署的網頁應用程式。模型下載後，在本地模型和應用程式可用的情況下，無須網絡連線也可分類；用於簿記或報税前，仍須核對輸出。
What I Built
LocalReceipt AI is a lightweight, privacy-focused expense classifier for freelancers and independent professionals.
It converts pasted receipt text into structured expense information, including category, amount, currency, date, and reason.
For example, the input:
Uber ride to client meeting, ₹420, 15 October 2026
produces a classification with category Transport, amount 420, currency INR, and date 2026-10-15.
The goal is to reduce repetitive expense organization and make bookkeeping preparation easier for freelancers and small business owners.
Demo
LocalReceipt AI currently runs locally on my Mac using Ollama and Flask.
Demo evidence: I tested the application with a sample receipt and confirmed that it returned the expected structured JSON output.
I have included a screenshot of the working application in this post.
This is a local prototype, not a publicly deployed web application.
Code
GitHub repository: https://github.com/Kishan-shah12/localreceipt-ai
The repository contains the source code for the Flask interface and its connection to the local AI model.
How I Built It
- Qwen2.5 3B: Open-weight language model used for expense classification.
- Ollama: Runs the model locally on my Mac.
- Python: Handles the application logic and local API requests.
- Flask: Provides the browser-based interface.
- JSON: Structures the AI's output into expense fields.
The browser sends receipt text to the local Flask application, which calls Ollama on the same machine and displays the result.
Why Does Open Innovation Matter?
Open-weight AI is central to this project because it allows expense classification to run locally instead of relying on a hosted AI API.
Privacy: Receipt text can remain on the user's machine rather than being sent to an external AI provider.
Cost control: Local inference avoids per-request hosted AI API charges after the model has been downloaded, although hardware and electricity still have costs.
Flexibility: The model and prompts can be changed to suit different freelance workflows without depending entirely on one provider's API pricing.
Offline capability: After the initial model download, classification can run without an internet connection, provided the local model and application are available.
Open innovation makes it practical to experiment with a privacy-conscious AI tool without requiring a paid inference API.
What I Learned
This prototype helped me explore how open-weight AI can power a focused, useful application on personal hardware.
The current version accepts pasted receipt text. Receipt image scanning, CSV export, and accounting integrations are possible future improvements. The output should be reviewed before being used for bookkeeping or tax reporting.
Prize Categories
Side Hustle — a practical AI tool designed to reduce repetitive expense organization for freelancers.
來源：DEV Community · dev.to