Hamdi Friji

- Built a LangGraph-based multi-agent system integrated with ServiceNow for automated incident management within an ITSM (IT Service Management) and AIOps (Artificial Intelligence for IT Operations) workflow: retrieves similar past incidents through semantic search in a RAG (Retrieval-Augmented Generation) pipeline, consults internal documentation and knowledge bases via agentic tool calling, queries server logs to extract relevant signals for RCA (Root Cause Analysis), and proposes ranked resolution steps to reduce MTTR (Mean Time to Resolution). - Deployed a two-stage Jina AI retrieval stack (jina-embeddings-v3 bi-encoder + jina-reranker-v2-base-multilingual cross-encoder) with Elasticsearch hybrid search (BM25 + dense vectors) on CPU-only infrastructure, using OpenVINO for inference optimization and quantization. - Benchmarked OpenVINO, ONNX (Open Neural Network Exchange) Runtime, and llama.cpp (GGUF quantization) to find the best latency–throughput trade-off on production CPU infrastructure. - Exposed services as REST APIs with FastAPI and deployed them as containerized microservices on Red Hat OpenShift (Kubernetes), following MLOps practices.

– Fine-tuned vision-language models (VLMs) to build document intelligence solutions for tasks such as information extraction and document classification. – Designed and built agentic systems to process clients’ emails. The main goal is to extract information, adapt behavior based on extracted information, and include a human-in-the-loop in cases of uncertainty. The agent achieved an automation rate of nearly 86% on received emails. – Enhanced the security of LLM-based pipelines by identifying and mitigating vulnerabilities to adversarial manip- ulations.

-I developed AI-based cybersecurity solutions, specializing in Graph Neural Networks (GNNs) applied to threat detection and model robustness. -Intrusion detection: Among the first researchers to apply GNNs to intrusion detection. I built systems that find complex malicious patterns in large datasets, designed for real-time threat detection and response. -IoT security: Designed and implemented multi-agent systems for distributed intrusion detection in IoT networks. I worked with interdisciplinary teams to deliver scalable solutions that adapt to changing environments. -Adversarial robustness: Researched adversarial attacks and built GNN-based models resistant to adversarial perturbations, which made the AI systems more reliable. -Generative AI security: Studied adversarial perturbations in the latent space of generative models to improve knowledge transfer while keeping generated content intact and secure. -Publications: Co-authored several scientific papers on AI security.

• Investigated a collaboration mechanism in a multi-agent system to overcome data-sharing limitations. • We investigated the feasibility of using adversarial perturbation techniques and GEN AI algorithm for secure information exchange among agents. • Enabled agents to manipulate latent space to generate data while retaining characteristics of the original dataset.

At Toshiba, I collaborated with a dynamic team to harness AI for groundbreaking advancements in cybersecurity. We developed a state-of-the-art 3-stage Intrusion Detection System, inspired by a streamlined version of the Lockheed Martin cyber kill chain. This system was specifically designed to identify complex, multi-step cyberattacks with greater precision. Our work is available online "Multi-stage Attack Detection and Prediction Using Graph Neural Networks: An IoT Feasibility Study".

During my work at Stevens Institute of Technology, I engaged in cutting-edge research and development of AI-driven solutions to tackle critical AI-related challenges. I focused on two primary projects: - Autonomous Driving Project: I proposed and developed innovative solutions to enhance an autonomous vehicle's car-following capability using Red Green Blue Depth (RGB-D) frames. This involved integrating the YOLOv3 object detector for identifying leading vehicles and obstacles, combined with a reinforcement learning algorithm to optimize navigation. This work significantly reduced the crash rate of autonomous vehicles by 5% and improved driving stability and smoothness, validated through a proof-of-concept demonstration in the CARLA simulator. - COVID-19 Pandemic Research: I contributed to a project that aimed to predict COVID-19 patient numbers across U.S. states using a mechanistic model, achieving high accuracy within a one-month range. Additionally, I worked on developing a graph-based framework for containing the spread of the pandemic, focusing on enforcing social distancing and calculating safe routes for pedestrians by identifying high-risk areas. Throughout these projects, I utilized my skills in Python and various AI techniques, enhancing my understanding of machine learning applications in real-world scenarios.