Σχεδίαση και ανάπτυξη ενός στρατηγικού εταίρου τεχνητής νοημοσύνης : μια ολοκληρωμένη αρχιτεκτονική βασισμένη σε δυναμικά ψηφιακά δίδυμα και μεγάλα γλωσσικά μοντέλα

Καμπέρης, Χρήστος Κωνσταντίνος

Master Thesis

Author

Καμπέρης, Χρήστος Κωνσταντίνος

Date

2026-02

Abstract

Despite their impressive capabilities, modern artificial intelligence systems typically function as passive assistants, where the user asks a question and the system provides a response. This thesis proposes and experimentally documents the transformation of this passive role into a true strategic partner, capable of actively analyzing data, simulating scenarios, and evaluating the reliability of its own outputs. The proposed architecture is based on the concept of a Dynamic Enterprise Digital Twin and is organized into four core pillars. These include Deep Knowledge through hybrid search, self-correcting RAG, and knowledge graphs; Reasoning with dynamic skill injection; Simulation via stochastic Monte Carlo analysis; and finally, Trust through the use of a composite reliability score. The coordination of these pillars is achieved via the Pillar Integration Context Protocol (PICP), a protocol that applies distributed systems principles to the coordination of artificial intelligence subsystems. For the evaluation of the system, seven experiments were conducted using a dataset comprising 544,629 real transactions. The data were extracted from the Enterprise Resource Planning (ERP) system of a Greek commercial enterprise, covering the period from 2004 to 2024. The primary findings were highly positive, with hybrid search achieving an nDCG@5 score of 0.926 (95% CI: [0.833–0.991]), while routing accuracy reached 80% with errors exclusively on ambiguous queries. The self-correcting mechanism achieved a 66.7% pass rate, with failure analysis revealing three distinct categories (cross-temporal synthesis, variable correlation, inferential analysis) that reflect the fundamental boundary between retrieval and reasoning. Concurrently, the Monte Carlo simulation demonstrated convergence at 500 samples with a retrospective coverage of 75%. Notably, the PICP protocol exhibited absolute resilience across eight different failure scenarios, with a negligible temporal overhead of less than 1 ms. Overall, this work offers a threefold contribution. At a theoretical level, it proposes a comprehensive framework for the integration of knowledge, reasoning, simulation, and trust. At a practical level, it introduces PICP as an original application of distributed coordination, while at an empirical level, it provides reproducible results through 459 automated tests (89.3% code coverage) and 9.291 lines of code.

Postgraduate Studies Programme

Πληροφοριακά Συστήματα και Υπηρεσίες

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων

Number of pages

163

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/19289

Collections

Τμήμα Ψηφιακών Συστημάτων

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα