Book Lister AI

The challenge

In the used-book trade the bottleneck isn't sales, it's data entry. Per book, staff used to spend 3 to 5 minutes photographing, transcribing (title, author, ISBN), researching prices, SEO-optimising, and uploading. At thousands of books per month that's enormous labour cost, before a single euro is earned.

The solution: hardware meets AI

An end-to-end pipeline that ties the physical scan process to multimodal AI and live APIs:

• Smart scanning. The book sits on a mat calibrated with ArUco markers. The webcam corrects perspective in real time, physically measures dimensions (for automatic shipping classes) and scans the barcode.
• AI data extraction. Two high-resolution scans (cover + back) go to Gemini 2.5 Flash. A strict JSON schema extracts title, author, publisher, year.
• Automatic pricing. Cross-checks against the Google Books API, queries the eBay Browse API for live competitor listings, and calculates a competitive price with profit-margin protection.
• Background upload. One operator click, a background worker pushes the listing live via the eBay Trading API while the next book is already being scanned.

Engineering highlights & fail-safe architecture

Absolute reliability was the core focus, the app runs in warehouse operations; downtime directly costs money:

• Trust-but-verify on AI data. Since LLMs occasionally hallucinate ISBNs, the architecture treats AI output as a hypothesis only. The ISBN is forcibly validated against the hardware barcode scan and a Google Books fuzzy match (thefuzz). Bad data is blocked before it corrupts the listing.
• Hybrid computer vision. Dual barcode-decoding system (ZBar for clean, zxing-cpp for damaged codes), maximum recognition rates even on old, scratched books.
• Thread-safe capture pipeline. To prevent Windows STATUS_HEAP_CORRUPTION crashes from competing camera restarts: strict VideoCapture ownership inside a dedicated, watchdog-monitored capture thread.
• Zero-touch database migrations. SQLite in WAL mode with automatic schema migration at app start. Every migration locked in by an explicit pytest suite. Updates roll without customer intervention.

The result

• +400% throughput. From 3–5 minutes per book to under 30 seconds.
• Cold start → first frame: 2–4 seconds.
• Scan → price: 4–6 seconds.
• ~6,450 LOC production code, locked in by ~2,720 LOC of unit tests (260+ pytest tests) + GitHub Actions CI.
• Deployment: 165 MB monolithic PyInstaller executable, double-click, done.