CIFRE - PhD Thesis – Root Cause Analysis and Impact Assessment with Generative AI in ITSM

Company Description

At EASYVISTA, we place great importance on the well-being of all our employees, which is why we were awarded the Great Place To Work® certification for 2022–2023.

Are you looking to join a fast-growing French software company that puts professional development and employee success at the heart of its strategy?

As a market leader in ITSM, EasyVista has recently acquired three technology companies specializing in ITOM, further strengthening and expanding our software offering.

Our goal is to grow globally and double our revenue over the next five years. To achieve this, we are investing in our people. Human values are central to our company culture, enabling our employees to thrive within expert, multidisciplinary teams.

Do you hold a Master 1 and have strong skills in data science, machine learning, and statistics?

Are you passionate about Big Data analytics?

Do you want to work in an Agile environment within an AI Lab team?

Are you looking for a PhD opportunity in computer science focused on causality and LLMs?

If you enjoy challenges and a collaborative culture, then don’t hesitate and join us!

Job Description

In an industrial and research-driven context, the candidate will investigate the application of generative AI models, particularly Large Language Models (LLMs), to tackle the complex tasks of Root Cause Analysis (RCA) and Impact Analysis . These tasks are critical in operational environments such as IT service management, incident response, and industrial systems monitoring, where understanding the origin and consequences of issues is essential for efficient resolution and continuous improvement.

The core objective of this thesis is twofold:

1. Evaluation of existing approaches to RCA and Impact Aanalysis using Retrieval-Augmented Generation (RAG) techniques, with a special focus on graph-structured information retrieval. While standard RAG methods typically rely on chunk-based retrieval from flat-text corpora, recent advancements such as Graph-RAG and Path-RAG [Chen et al., 2024] propose structuring the knowledge base as a graph to better reflect semantic dependencies and relational knowledge. These approaches enable more precise and logically coherent responses by guiding LLMs along meaningful information paths. The candidate will assess the precision, exhaustiveness, and production viability of these methods in real-world environments, paying close attention to redundancy control and retrieval efficiency.
2. Proposition of a novel hybrid framework that advances the current state of the art by combining causal reasoning capabilities with graph-augmented retrieval. Traditional LLMs are primarily correlation-based and may fail to distinguish spurious correlations from true causal relationships. Integrating causality into LLMs can significantly improve their reliability, reduce hallucinations, and enhance their ability to perform accurate RCA. Recent studies [Wu et al., 2024] have underlined the limitations of prompt-based causal reasoning and the need to embed causality throughout the LLM lifecycle—from training to inference. The candidate will explore how causality-aware architectures or training strategies can be combined with RAG or graph-based prompting to improve both interpretability and robustness.

This PhD project will involve a combination of machine learning research, natural language processing, and knowledge graph engineering, with concrete applications in industrial settings. The expected contributions include:

• A benchmark comparison of existing RAG and Path-RAG approaches for RCA/IA tasks.
• An in-depth evaluation of LLMs' causal reasoning abilities in operational contexts.
• A novel methodological framework that integrates causality and graph-based retrieval into generative pipelines.

References:

Chen, B., Guo, Z., Yang, Z., et al. PathRAG: Pruning Graph-based Retrieval Augmented Generation with Relational Paths. 2024.

Wu, A., Kuang, K., Zhu, M., et al. Causality for Large Language Models. 2024.

Qualifications

With a background in computer engineering, specialized in data science, and significant experience in distributed application architectures with high constraints and large-scale data processing.

• Strong command of Machine Learning concepts
• Good understanding of statistical and mathematical models
• Design and implementation of new analytical models
• Familiarity with NLP (Natural Language Processing) models
• Development of machine learning algorithms in Python
• Agile methodologies and industrialization practices: unit testing, Git, etc.

Additional Information

These tasks will be carried out in an Agile environment that has been in place for several years, with a strong focus on team collaboration and close cooperation with various stakeholders involved in the product. A results-oriented mindset, initiative, and the ability to make proactive suggestions will be key assets for the success of this internship.