📄 View CV Online | 📥 Download PDF

• LinkedIn: pepealessioo
• GitHub: pepealessio

AI Research Engineer with expertise in Generative AI, Adversarial Robustness, and Edge Computing. Proven track record of taking systems from research prototypes to production-grade deployment in safety-critical environments. Analytical mindset with a focus on critical thinking, problem-solving, and attention to detail.

• June 2024 – present
• Rome (IT)
• Focused on Natural Language Processing (NLP) and adversarial machine learning, leading modular AI framework development, and improving robustness and safety of AI systems in production and operational environments.
• Led development of a low-code agentic AI framework (ZeroMQ-based async microservices) supporting multi-modal LLMs, RAG, MCP and ReAct agents, scaling it from proof-of-concept to a productized solution.
• Deployed object detection models on edge hardware during the NASR25 military exercise, achieving [email protected]=0.92 on small objects in aerial imagery, helping secure ~€30M in potential follow-on contracts.
• Implemented 10+ adversarial ML attacks, including gradient-free black-box methods, evolving a proof-of-concept into a ~€2M project acquired within the Global Combat Air Programme (GCAP).
• Built physical adversarial attack pipelines for object detection in real-world environments, with patches reducing [email protected] by up to ~70 pts and extending from legacy YOLOv2/v3 to latest-generation YOLO (e.g., YOLOv12).

• Sept 2023 – June 2024
• Rome (IT)
• Worked on computer vision and deep learning systems for smart-city applications, focusing on people/vehicle detection and traffic analysis. Deployed solutions on edge devices embedded in smart cameras.
• Trained models for Automatic License Plate Recognition (ALPR), boosted accuracy (mAP) by 46%.
• Optimized C++ camera backend, reducing CPU usage by 30% through asynchronous processing.
• Collected, cleaned, and labeled diverse real-world datasets to improve model generalization.

• Mar 2023 – June 2023
• Fisciano (IT)
• Assisted OS and CPU Architecture students, providing support for MIPS assembly programming and design exercises.

• Sept 2022 – Dec 2022
• Enschede (NL)
• Joint internship with the University of Salerno, exploring video generation with controllable emotion using GANs.
• Designed a novel GAN loss that improved emotion accuracy by 40% while reducing identity drift by ~2%.
• Conducted human evaluation studies with 70% preference over baseline models.

• Jan 2023
• Fisciano (IT)
• GPA: 3.98/4.00, graduated with Honors (Transcript)

• July 2020
• Fisciano (IT)
• GPA: 3.85/4.00, graduated with Honors (Transcript)

• Programming: Python, C++, C, Rust, Java, SQL
• AI and ML Tools: PyTorch, TensorFlow, Hugging Face, vLLM, llama.cpp, LangChain, LangGraph, OpenCV, MCP, ReAct
• Edge AI: TensorFlow Lite, ONNX, Coral TPU, GStreamer, Async Microservices
• Backend & Tools: FastAPI, Flask, Qdrant (Vector DB), Git, Unix/Linux, ZeroMQ
• Languages: English (fluent - C1), Italian (native)

• Nov 2025
• Leonardo
• Honorable Mention, “Best Development 2024” for Fproto, a low-code agentic AI framework for rapid pipeline prototyping.

• Mar 2023
• University of Salerno
• Recognized among the top students of the previous 5 years in the DIEM department.

• June 2025
• IBM

• Jan 2025
• DeepLearning.AI

• Jan 2025
• University of Alberta

• Trained a face identification system based on ResNet50, achieving up to 99.5% accuracy.
• Evaluated robustness (FGSM, BIM, PGD) showing that small perturbations (~0.64% of pixels) can fool both classifiers.
• Designed detectors with 93–97% correctly flagged attacks (TP) while misclassifying <7% (FN) of clean samples.

• Implemented three parallelized versions of the Counting Sort Algorithm using OpenMP, MPI, and CUDA, achieving respectively 2x, 3x, and 20+x speedup compared to the sequential version.

• Built a real-time behavioural planner and perception pipeline (dual RGB + depth + semantic segmentation + YOLOv5-s) for a self-driving car in CARLA, correctly handling ~97% of 90+ safety-critical events across 20 urban scenarios.

• Designed a shopping bot application for the Pepper robot, integrating with its core and speech modules.
• Developed voice-based person re-identification functionality, enabling the robot to recognize returning customers.