Lorenzo Germano
AI Analyst @ Accenture
Based in Milan, ITCitizenship: IT,
EU
Tempus est vir honestus
Passionate about making AI accessible. At Accenture, I've led teams in developing AI agents, built custom agentic architectures, and created demos that showcase cutting-edge AI capabilities. My research background includes novel frameworks for 6-DoF object tracking and cross-modal attention mechanisms. I hold a M.Sc. in Mathematics in Data Science from TU Munich, and a B.Sc. in Physical Engineering from PoliTo.
Multimodal AIComputer VisionAI Applications
Portrait credits
Latest Projects
2025
Reinventing national service delivery
Case study on how Accenture helped Poste Italiane transform its national service delivery platform
2025
Pupil Detection from Event Frame Sequence
Experiment in event-based computer vision: A novel approach for pupil detection using event camera data, leveraging temporal information and intensity changes for robust eye tracking.
This project explores pupil detection using event camera data, a novel approach in computer vision that leverages the temporal information inherent in event-based sensors. Event cameras, unlike traditional frame-based cameras, capture intensity changes at each pixel independently and asynchronously, providing high temporal resolution and low latency. The project investigates how this unique data modality can be harnessed for robust pupil detection and eye tracking applications. The methodology involves processing event frame sequences to detect pupil boundaries by analyzing intensity changes over time, particularly focusing on the contrast between the pupil (darker region) and the surrounding iris (lighter region). The research addresses challenges in event-based vision including noise filtering, temporal consistency, and real-time processing requirements. Methodology includes: (1) use of event-based preprocessing techniques for noise reduction, (2) implementation of temporal filtering algorithms to enhance pupil boundary detection, (3) evaluation of the approach's robustness under varying lighting conditions and eye movements. The project has potential applications in human-computer interaction, assistive technologies, and medical imaging where precise eye tracking is essential.
2025
P2ROS: Prompt to Robotic Operating System
React-based demo for an AI-powered platform concept that would train robotic arms through imitation learning. All back-end functionality is simulated - you can click through the complete process to reach the final training dashboard.
The P2ROS demo represents a React-based prototype for an AI-powered robotic automation platform concept designed to make robotics more accessible as robotic arm prices drop and TOF sensors become commercially available. The demo targets small and medium businesses, enabling them to evaluate the benefits of robotic systems even without technical expertise. The concept leverages advances in imitation learning, where robots would learn complex tasks from human demonstrations. Key innovations in the concept include integration of consumer-grade 3D scanning (iPhone/iPad LiDAR), computer vision for motion extraction, and AI-powered code generation that converts natural language descriptions into executable robot control programs. The simulated workflow demonstrates how users would provide detailed task descriptions, capture 3D workspace environments, and record demonstration videos, which AI would process to train industrial robotic arms. All back-end functionality is simulated, but you can click through the complete process from initial setup to the final training dashboard. The platform concept addresses the convergence of affordable robotic hardware, breakthrough AI learning algorithms, consumer 3D scanning technology, and exponential AI capabilities, representing how robotic automation could become accessible to businesses of all sizes through an intuitive, demonstration-based training paradigm.
2024
Demo XR Tarantismo
Research prototype for immersive cultural heritage preservation using extended reality (XR), exploring methodologies for blending digital and physical experiences.
This project proposal follows a conversation with Barbara Perrone on ideas about how to present the ritual of Tarantism to visitors through a contemporary medium. Inspired by a visit to the Museo Civico Pietro Cavati, where I studied relevant materials, I designed a research prototype that explores cultural heritage preservation through extended reality technologies. The conceptual framework, designed for Meta Quest 3 or Pro headsets, investigates how the Tarantismo cultural phenomenon from southern Italy can be experienced through multi-layered spatial computing. The demo models how visitors could navigate the Chapel of San Paolo delle Tarantate in Galatina while accessing multimedia content through exhibition panels and dimensional portals. This approach potentially addresses key challenges in cultural preservation, including engagement with intangible heritage elements such as crowds, and spatial context recreation. Future development would require 360° recording of multimedia content. The prototype represents an interdisciplinary research approach connecting technology, anthropology, and spatial computing.
The video and texts are purely illustrative. The demo was rendered using the Meta XR Simulator.
2024
Palazzo Elefante della Torre
Website design and development for a luxury vacation rental property housed in an 18th century baroque palace in Galatina, Italy.
For this project, I designed and deployed a website for Palazzo dell'Elefante della Torre, a luxury vacation rental business housed in an 18th century baroque palace in the historic center of Galatina, Italy. The website was built on the Lodgify vacation rental platform, which I customized using HTML and CSS and a copilot to achieve what I consider a unique, premium aesthetic that aligns with the luxury positioning of the property. Beyond the technical implementation, the project required a deep understanding of legal and commercial aspects specific to the hospitality industry, including Italian tourism regulations, booking terms and conditions, cancellation policies, and privacy compliance with GDPR. I created an elegant, responsive design that highlights both the palazzo's baroque heritage and its contemporary amenities, including rooms featuring mid-century Italian design pieces by notable designers like Giò Ponti and Ettore Sottsass. The implementation includes photography displays, and a seamless integration with Lodgify's booking engine.
2023
Featured
3D Human-Object Interaction in Video
Research in computer vision: A novel framework for 6-DoF object tracking in RGB video using cross-attention mechanisms, named H2O-CA.
This research project proposes a novel framework for 6-DoF object tracking in monocular RGB video. The H2O-CA framework investigates a sequence-to-sequence approach that leverages parametric human body modeling and applies a cross-modal attention mechanism to correlate human pose with object positioning. The contributions include: (1) a mathematical model for attention-based relationship modeling between human and object movement, (2) an algorithmic approach to handling partial or complete object occlusion, and (3) a sliding-window methodology for temporal consistency in tracking. The research methodology involved creating a conceptual implementation using PyTorch. This work has potential applications in augmented reality, human-robot interaction, and activity recognition systems where understanding spatial relationships between humans and objects is crucial. The research extends previous work on transformer architectures for computer vision.
2021
HeimGuard Safety AI Think Tank
Conceptualization of AI-based home safety systems combining computer vision, anomaly detection, and IoT sensors for preventive risk identification, now being developed by commercial entities.
HeimGuard was a forward-thinking conceptual project developed in collaboration with Charleen and Kasper that explored the potential of AI and IoT systems for enhancing home safety. This think tank initiative addressed the significant issue of household accidents by conceptualizing an integrated safety monitoring system that could identify and prevent common domestic hazards. Our German-language presentation outlined a comprehensive AI-driven safety system that utilized computer vision and smart sensors to monitor for five primary household risks: falls (particularly on stairs), cuts (especially in kitchen environments), burns, animal-related incidents, and accidental poisonings. The proposed system featured a central processing unit that analyzed camera feeds to detect potentially dangerous situations, with interactive capabilities through speakers and displays to provide warnings and guidance. Key innovations included fall detection with automated emergency response, kitchen safety monitoring with knife handling guidance, pet behavior analysis drawing from animal psychology, and poisoning prevention through monitoring of chemical storage. The system emphasized privacy-conscious local processing and edge computing rather than cloud-based solutions. Several commercial entities have developed remarkably similar technologies, including Aivilon's video analytics platform and Sparknets' Neurospot CCTV AI system. These commercial implementations demonstrate the real-world value of AI-powered safety monitoring systems, particularly for vulnerable populations such as the elderly and children.
2020
A Simple Digital Filter
Design of a digital filtering techniques in VHDL, optimizing performance through power-of-two coefficient design.
This project presents the design of a digital filter implementation in hardware, accessible on GitHub. The focus was on processing efficiency for a system handling 1024 8-bit samples using the transfer function Y(n) = -0.5X(n) - 2X(n-1) + 4X(n-2) + 0.25X(n-3). The methodology employed power-of-two coefficients by replacing multiplier units with shift operations. Key aspects: (1) developing an overflow detection mechanism for 2's complement arithmetic, (2) designing a Moore-type state machine controller for process management, and (3) creating an efficient datapath architecture using shift registers. The methodology included analysis, simulation-based validation, and performance benchmarking. This research has applications in resource-constrained embedded systems, IoT devices, and real-time signal processing hardware where computational efficiency is critical.
Work Experience
My journey bridges research and industry, with a focus on building innovative AI applications.
Curriculum Vitae
Education
Claude Certified Architect: FoundationsNEW
Present- Designed agentic loops and multi-agent systems using coordinator-subagent patterns, with task decomposition, session state management, and workflow enforcement
- Defined tool interfaces with clear boundaries and structured error responses, integrating MCP servers and distributing tools across agents
- Configured CLAUDE.md hierarchies, custom slash commands, and path-specific rules; integrated Claude Code into CI/CD pipelines
- Engineered prompts with explicit success criteria and few-shot examples, enforcing structured output via JSON schemas with validation and retry loops
- Preserved context across long interactions, designed escalation and error-propagation patterns, and applied confidence calibration in multi-agent reliability scenarios
M.Sc. Mathematics in Data Science
March 2021 — March 2024TU Munich — 2024 QS World University Ranking by Subject - CS & Inf.: 29
- Strong mathematical fundamentals in probability, optimization, and linear algebra
- Developed and implemented machine learning solutions for computer vision tasks (multi-object tracker), recommendation systems, robustness certification, NLP tasks
- Implemented appearance-based tracking using a Re-ID network on the MOT16 dataset achieving high accuracy in detection, segmentation, and tracking
- Improved IDF1 score of a ReIDTracker by implementing a customized Message Passing Network on bipartite graphs
- Used Google Cloud Platform for deploying, scaling, and maintaining pipelines
- Implemented two normalizing flow (NF) layers, stacked transformations, and trained the model with maximum likelihood on 3 datasets
- Crafted adversarial examples on MNIST on a CNN, and certified robustness via the principle of randomized smoothing
- Implemented Skip-Gram model for obtaining vector representations for words from a text, achieving a training accuracy of 80%
- Conducted spectral clustering on Yelp! social network to identify groups of users with similar preferences
- Achieved 85% accuracy in semi-supervised node classification using Graph Neural Networks on Cora ML citation network
- Implemented logistic regression with gradient descent and L2 regularization using only built-in Python and NumPy functions
- Implemented SVM dual problem using CVXOPT library for Quadratic Programming
- Performed Principal Component Analysis (PCA) on dataset by calculating covariance matrix, eigenvalues, and eigenvectors
- Developed a restaurant recommendation system using a latent factor model with regularization
- Implemented image compression algorithm using k-means clustering
- Developed an autoencoder using PyTorch Lightning to pre-train on 60,000 unlabeled images from the MNIST dataset
- Built and trained a regression CNN model for facial keypoint detection with less than 5 million parameters
- Designed and implemented a PyTorch model for image segmentation
- Implemented sentiment analysis model using RNNs/LSTMs to predict the sentiment of movie reviews
- Implemented and experimented with Recurrent Neural Networks (RNN) for various applications
- Processed large datasets using various libraries and frameworks (query optimization and page-ranking in C++, advanced SQL query decorrelation)
- Optimized a join on the TPC-H dataset (MMAP, blockwise search, parallelism) in C++
- Developed an Apache Spark program to analyze taxi rides in New York
- Implemented a KNN algorithm for a directed graph in C++ and made it efficient
- Ethics management in data analysis and data protection processes
Guest Student
August 2022 — January 2023Denmark TU — 2022 QS World University Rankings by Subject - Eng. & Tech.: #60
- Courses completed: Partial Differential Equations, Image Analysis, Machine Learning Operations
B.Sc. Physical Engineering
October 2017 — September 2020Politecnico di Torino — 2020 QS World University Ranking by Subject - Eng. & Tech.: #38
- Skills in advanced and applied physics (quantum and statistical mechanics, complex systems physics, solid state physics, nuclear physics)
- Classical treatment of the interaction of electromagnetic radiation with matter, introduction to surface plasmons, photonic crystals
- Scattering of EM radiation (Raman effect and spectroscopy), light-detecting devices, spontaneous and stimulated emission
- Advanced optical lithography (standard photolithography, stereo photolithography, and 2-photon processes, CVD, ALD, etching)
- Basic engineering skills in the field of information engineering (Electrical Engineering, Electronics, Electromagnetic Fields)
- Technologies for producing vacuum and preparing environments
- Basic elements of cryogenics
- Technologies for growing materials and producing thin films, surface modification techniques
- Microscopy techniques
- Skills of modern information engineering (Introduction to Quantum Computing Algorithms, OpenQASM)
Competitions
AI S&Cale Up for Analyst Hackathon
March 2025- 1st Place Award in the Third Edition of the AI S&Cale Up for Analyst challenge
- Led a team of 12 analysts to create the best Agent against 11 competing teams in the "Create Your Agents with Amethyst Studio" challenge
Football: Can you guess the winner? Challenge
February 2024Qube Research & Technologies — Challenge Data
- Developed a classification model to predict the outcome of a match from historical data at the player, team, and league levels
- Score: 48.86 / 49.32
Analytics Cup (DS Competition)
March 2021- Developed a classification model to predict the success of sales offers for smart infrastructure using Scikit-learn
- Challenge: unbalanced dataset
- Score: 65.05 / 68.38
Athletics National Championship
September 2014- 8th place in the Italian championship of Decathlon for the category Allievi
Honors
Merit-based Scholarship
2018 — 2020National Register of Excellence
2017Ministry of Education, University and Research — Albo Nazionale delle Eccellenze
- Listed in the National Register of Excellence for achieving a perfect score of 100 with honors in the state exam
- School code: TOPS01000G, Academic Year: 2016/2017
Conferences and Seminars
3D Human-Object Interaction in Video
June 2024Research Presentation, Host: Visual And Multimodal Applied Learning Laboratory (VANDAL), Politecnico di Torino
- Presented novel research on 6-DoF object tracking in RGB video using cross-attention mechanisms to model human-object interactions
- Demonstrated how the framework leverages human pose to improve object pose estimation, especially in scenarios with occlusions
Ethics of Responsibility: An Introduction to Applied Ethics
September 2022Seminar, Organizer: PD. Dr. Jörg Wernecke, TUM School of Social Sciences and Technology
- Presented on "Climate Change as Problem of Many Hands" (30'), discussing how responsibility-as-blameworthiness, responsibility-as-accountability, and responsibility-as-virtue-and-obligation circa climate change share inside a collective
Randomized Algorithms
April 2021Advanced Seminar, Organizer: Prof. Dr. Robert König, TUM School of Computation, Information and Technology
- Presented on Markov chains and random walks, sorting and median estimation, fundamentals of sorting (30'/each) with a proof-based approach to an audience of twenty master's students
Volunteering
Helix Aquatics fundraiser
May 2016- Collaborated in a fundraising event for the Helix Charter HS Aquatics team with over 200 attendees, raising more than $30,000
Boys and Girls Club of Carlsbad
November 2015- Volunteered in an after-school recreational center
- Helped kids with homework, organized activities, and decorated the building