How AI Turned My Chaotic Home Inventory into an Organized System

Problem Overview

After moving, the author faced four main issues in household goods management: wasted storage space due to expired items, food waste from missing expiry tracking, difficulty locating items without proper labeling, and duplicate purchases caused by unclear inventory.

Solution Overview

During the Spring Festival the author built a home‑goods management application using the AI IDEs Cursor and Trae (Claude‑3.5‑Sonnet and deepseek‑r1). The goal was to digitize inventory, improve efficiency, and provide convenient reminders.

Main Features

Inbound Management : Users record name, category, expiry date, production date, etc. Initially planned barcode scanning with GPC classification, but low efficiency for items without barcodes led to AI‑based smart recognition.

Outbound Management : Records usage or discard events and updates inventory in real time. Supports both barcode scanning and image‑based search for items lacking barcodes.

Near‑Expiry Reminder : Displays items approaching expiry; future integration with HomeAssistant for push notifications.

Architecture

Implementation Details

Barcode Recognition

The app uses Alibaba Cloud’s barcode query API for fast, accurate product information. Development was driven by providing Curl examples to Cursor.

Product Image Recognition

Prompt engineering was crucial. The author split brand and category recognition into a chain‑of‑thought prompt and used color as a fallback when the brand was unclear. The chosen model was Doubao-1.5-vision-pro-32k.

# Role
You are a professional image‑recognition expert...
(omitted for brevity)

Standardized Classification

To avoid custom taxonomy issues, the international GPC (Global Product Classification) system with four levels and over 6,000 categories was adopted. Initial RAG retrieval using product names performed poorly; adding a generation step to filter noisy words improved matching. The generation step used Doubao-1.5-lite-32k for speed.

# Role
You are a professional product‑classification expert...
(omitted for brevity)

Expiry and Production Date Extraction

OCR models were used to read dates, then converted to days (e.g., 1 year = 365 days). The same Doubao-1.5-vision-pro-32k model was employed.

# Role
You are a professional image‑recognition expert for shelf‑life extraction...
(omitted for brevity)

Image‑Based Search

For items without barcodes, images are vectorized and stored. At checkout, the query image is vectorized and matched against stored vectors using Alibaba Cloud’s DashVector service. Models used: Doubao-embedding-vision for embedding and DashVector for retrieval.

Experience Summary

Cursor vs. Trae

Code Generation Quality : Both use Claude‑3.5‑Sonnet, resulting in similar quality.

Generation Speed : Cursor Pro was noticeably faster; Trae suffered stability issues and frequent errors on large outputs.

Multimodal Input

Both support image input when the model permits.

IDE Capabilities

Web Preview : Trae offers this feature, but its built‑in browser cannot access camera APIs, making it unusable for this app.

Agent Ability : Trae’s Builder mode includes agent functionality by default; Cursor requires manual selection of Composer or normal mode. In practice, both agents performed similarly, automatically executing CLI commands and handling lint errors.

Claude‑3.5‑Sonnet vs. deepseek‑r1

Efficiency : deepseek spends more time on “think” steps before code modification, reducing overall speed compared to Claude.

Professional Handling of API Keys : Claude stores keys in environment variables and creates .env files, whereas deepseek writes keys directly in code, making Claude’s approach more professional.

Source Code

https://github.com/jsceoz/goo