Analysis, Evaluation and Benchmarking of Complex Agentic AI Systems
SCK CEN aims to integrate Artificial Intelligence into its activities in a secure way, by deploying AI services locally within its own premises. Given the sensitive nature of the nuclear sector, security and data confidentiality are major concerns: information must remain under SCK CEN’s control at all times, which rules out most commercial, cloud-based AI products and evaluation platforms offered by third-party providers. As a consequence, SCK CEN is developing in-house the tools required to build, operate and assess AI services based on Large Language Models (LLMs). Among these, agentic systems, in which several LLM-based agents cooperate, call external tools and exchange information to solve complex tasks, are particularly promising. They are also notoriously difficult to evaluate: errors may propagate from one agent to another, hallucinations may go unnoticed, and the impact of a given design choice or parameter on both quality and cost is hard to quantify. The present internship primarily focuses on addressing this gap.
The main objective of this internship is the design and implementation of a tool for the analysis, evaluation and benchmarking of complex agentic AI systems, capable of diagnosing such a system based on a wide range of criteria. Approaches to be investigated include, among others, (i) criteria-based evaluation of the system outputs (e.g., correctness, relevance, completeness, latency and cost); (ii) LLM-as-a-judge techniques for automated assessment; (iii) analysis of error and hallucination rates, both at system level and at the level of individual agents; (iv) systematic tuning of the system parameters (models, prompts, retrieval settings, agent configuration) to study their influence on the quality-vs-costs trade-off; and (v) human-in-the-loop evaluation to validate and calibrate the automated metrics. This list is not exhaustive and is expected to evolve as the work progresses; the tool shall therefore be designed in a modular and extensible way, so that new evaluation methods can easily be integrated. In line with SCK CEN’s security requirements, the tool shall run entirely on-premises and rely on locally hosted models.
Secondly, the tool will be validated on representative agentic use cases in order to assess its relevance and to produce a first comparative benchmark. One of the main expected outcomes of this internship is to turn this prototypical evaluation system into a practical asset for the design engineers of the NTE group at SCK CEN, particularly in the context of the LEANDREA project. The results shall be documented so as to support future design decisions.
The ideal candidate for this position should be familiar with programming in Python as well as with Python frameworks for the creation of Generative AI services (e.g., LangChain, LangGraph). Knowledge of LLM evaluation methods and basic statistics will be beneficial.
This work will be supervised by Dr. Matteo Greco (mentor) and co-supervised by Vincenzo De Florio (co-mentor).