Receipto is a lightweight, offline-first smart receipt reader designed for Android devices. It leverages a combination of classical computer vision and a localized TensorFlow Lite Key Information Extraction (KIE) engine to extract structured financial data such as Store Name, Date, Amounts, and Tax directly on-device. This ensures user privacy as no cloud APIs are involved, making it a fully offline solution.

Core Features

• On-Device Processing: Utilizes ML Kit Text Recognition and custom KIE/heuristic algorithms to extract data without an internet connection.
• Layout-Aware OCR Parsing: Groups bounding boxes from raw text to reconstruct logical lines and accurately pair prices to items, even on crinkled receipts.
• Offline Persistence: Automatically saves extracted data to a local Room Database for seamless retrieval and history.
• Fallback Heuristics: Uses deterministic spatial fallback rules to handle chaotic layouts when ML prediction confidence is low.

How it Works

The pipeline is designed to be deterministic and layout-aware, prioritizing predictable results over purely black-box ML predictions. It includes image capture and preprocessing using OpenCV, raw token extraction with Google's ML Kit, and key information extraction using TensorFlow Lite. Spatial heuristics are applied when necessary to ensure accurate data extraction.

Architecture

The project adheres to Clean Architecture principles, with a clear separation of concerns between the UI layer, domain logic, and data persistence. The UI is built using Jetpack Compose, while data is managed using Room Database.