A.B.C. – A.I. Book Cataloguer: from concept to working prototype

28 February 2026   |   News   |   Tags: AI, book scanner, books

This article follows up on our previous post: Axelera AI Smarter Spaces Project Challenge, where we announced our participation in the challenge.

From concept to working prototype: our journey through the Axelera Project Challenge.

The project

A.B.C. (“A.I. Book Cataloguer”) is an edge-AI system for automatically cataloguing books, developed by us at Denovo, an Italian startup, as part of the Axelera Project Challenge 27. The project was first announced on the Axelera community on December 3, 2025, and reached a fully working prototype on February 25, 2026, with final documentation published on February 28, 2026.

The core idea is straightforward but powerful: we replace the manual, slow, repetitive and error-prone process of book cataloguing with a smart tabletop station that scans a book cover in real time and extracts title, author and publisher automatically, without requiring any cloud connectivity.

The team

We are an innovative startup in the field of AI-driven mechatronics, a small international team made up of:

• Two Italian engineers
• One Polish project manager

More information about the team is available on the Team page.

A.B.C. is our first real edge-AI project, brought to life thanks to the hardware kit provided by Axelera as part of the challenge.

The problem A.B.C. aims to solve

Digitising book collections has long been a slow and labour-intensive process. Operators must read each book, manually enter data and correct errors – a workflow that slows down operations and undermines metadata quality.

We designed A.B.C. to eliminate this bottleneck by automating information extraction and introducing a natural, gesture-based correction mechanism. This makes the entire process faster, more accurate and far more intuitive.

Potential applications include libraries, second-hand bookshops and comic book stores, where continuously cataloguing ever-changing stock is a real and costly challenge.

The hardware

The system is built around the kit provided by Axelera, which arrived from the Netherlands in late December 2025. The components used in the final setup are:

• SBC: Orange Pi 5 Plus
• NPU accelerator: Axelera Metis M.2
• Camera: Sonoff CAM S2
• Local network: TP-Link router with direct Ethernet connection
• Camera mount: ring light with adjustable arm
• Levelling: dual-bubble level attached to the camera
• Storage: SD card
• Display: monitor

Software and AI pipeline

After several iterations, we settled on the final solution based on PP-OCR (PaddleOCR v3 Latin) in hybrid mode, after discarding EasyOCR, Tesseract and several LLMs (which proved too “creative”, fabricating data instead of correcting it).

The complete pipeline involves:

1. RTSP camera acquisition and loading-area crop via OpenCV
2. Book detection with YOLOv8l running on the Metis NPU
3. Hybrid OCR sub-pipeline: text-block detection on Metis and actual OCR acquisition on CPU with a newer, more accurate model
4. Fuzzy matching against a local database (12 GB Open Library dump)
5. On-screen result and wait for the operator’s gesture

Operator interaction and UX

The operator interacts with the system in a completely hands-free, gesture-based way:

• Placing a book in the loading area automatically starts the process
• Sliding the book away confirms the acquisition – a large green check appears and data is saved to a local CSV file
• Showing crossed fingers (an X shape) under the camera triggers a discard – a large red X appears
• Both actions feature a time bar, giving the operator a window to interrupt the action

In the final phase, we added a QR code command system that lets the operator re-trigger calibration, toggle diagnostics and handle a clean shutdown of the Orange Pi.

Project timeline

• Dec 3, 2025: first post published on the Axelera community
• Dec 20, 2025: hardware kit received from the Netherlands
• Jan 12, 2026: full unboxing, Voyager SDK installation, first test with “inference.py”
• Feb 1, 2026: physical setup, Open Library DB integration, code published on GitHub
• Feb 23, 2026: complete pipeline, switch to PaddleOCR v3, final debugging
• Feb 25, 2026: working prototype completed
• Feb 28, 2026: final documentation and verified physical setup

Axelera features the project

On April 10, 2026 Axelera publicly featured our project on their LinkedIn profile: see the post (opens in a new tab).

Results and outlook

We completed the system successfully within the challenge timeframe. The final prototype operates entirely offline (with the sole exception of the Wi-Fi link between camera and local access point) and demonstrates the viability of an automatic cataloguing system running on low-cost edge hardware.

All code, documentation, instructions, QR codes, images and the Orange Pi “.stl” support bracket are freely available in the project’s GitHub repository: github.com/denovo-it/abc (opens in a new tab).

Project developed by us at Denovo as part of the Axelera Project Challenge 27 – December 2025 / March 2026.

ABC (AI Book Cataloguer) is one of the public projects from our artificial intelligence stream, developed in‑house alongside mechatronics, blockchain and cybersecurity.