Shabti KI Engine (SKE) — A Domain-Specialised AI System for the Scholarly Analysis of Ancient Egyptian Shabtis

Abstract

The Shabti KI Engine (SKE) is a domain-specialised artificial-intelligence system for the scholarly analysis of ancient Egyptian shabtis, developed at the private research collection Sammlung Beckers (Aachen). SKE combines a fine-tuned vision-language model (`qwen3-vl-shabti:32b`, LoRA-adapted on period and material tasks) with a DINOv2-based linear-probe classifier for chronological attribution, a citation-grounded iconographic-feature module, a visual-similarity search on DINOv2 embeddings, a fake-detection pipeline operating on original image bytes, and a first-phase inscription-analysis module with confidence gating. All modules operate around a shared analysis kernel serving both internal use and, via a strict peer-firewall, external experts through a peer-review interface. The system runs on-premise on a two-host infrastructure (a Windows workstation with two NVIDIA RTX 5090 GPUs hosting VLM and embedding services, and a Linux Intel NUC hosting the WebApp, MongoDB, and ChromaDB), with a rigorous hold-out evaluation discipline preventing self-match contamination. The core scientific result at the time of writing is the DINOv2 linear-probe period classifier, which reaches an accuracy of 0.803 and macro-AUC of 0.946 on a hold-out of 67 objects (15 % of a book-attested corpus of 440 objects; zero overlap with the training partition), thereby resolving the persistent NK/TIP collapse of the holistic VLM approach (baseline 0.379) across all periods represented in the hold-out. This paper documents the current architecture, module status, evaluation methodology (including class distribution, per-class accuracy, and known caveats), hardware infrastructure, and integration with the Deep Publication workflow of the Beckers Collection. It is published on Zenodo and Academia.edu as an open concept paper inviting scholarly collaboration, dataset exchange, and methodological critique.