Page Not Found
Page not found.
Sitemap
A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.
Pages
About me
About me
Professional Experience
Professional experience of Paolo Di Francesco — AWS Solutions Architect, cybersecurity, and academic research.
Publications
Academic publications by Paolo Di Francesco in IEEE, ACM, and other venues.
Teaching
Teaching experience — TU Wien external lecturer and Trinity College Dublin teaching assistant.
Posts
Simplify ModelOps with Amazon SageMaker AI Projects using Amazon S3-based templates
Managing ModelOps workflows can be complex and time-consuming. Amazon SageMaker AI Projects now offers an easier path with Amazon S3-based templates. With this new capability, you can store AWS CloudFormation templates directly in Amazon S3 and manage their entire lifecycle using familiar S3 features such as versioning, lifecycle policies, and cross-region replication.
Track LLM model evaluation using Amazon SageMaker managed MLflow and FMEval
In this post, we show how to use FMEval and Amazon SageMaker to programmatically evaluate LLMs. FMEval is an open source LLM evaluation library, designed to provide data scientists and machine learning (ML) engineers with a code-first experience to evaluate LLMs for various aspects, including accuracy, toxicity, fairness, robustness, and efficiency.
Create a Generative AI Gateway to allow secure and compliant consumption of foundation models
In the rapidly evolving world of AI and machine learning, foundation models have shown tremendous potential for driving innovation. However, as organizations increasingly harness the power of FMs, concerns surrounding data privacy, security, added cost, and compliance have become paramount. In this post, we propose a Generative AI Gateway as a platform for an enterprise to allow secure access to foundation models for rapid innovation.
Secure MLflow in AWS Fine-grained access control with AWS native services
MLflow and Amazon SageMaker are two of many tools on the market to help data scientists to implement end-to-end Machine Learning workloads. SageMaker offers the possibility to run these workloads fully end-to-end on its own ecosystem as it has been designed to solve some of the common challenges that are peculiar to ML lifecycle workloads. Nonetheless, one of the great traits of the SageMaker ecosystem is also its flexibility and openess to integrate with other tools. Today, we ultimately want to show how you can securely integrate SageMaker with MLflow using native AWS services to enable access control on the open-source version of MLflow.
Organize your machine learning journey with Amazon SageMaker Experiments and Amazon SageMaker Pipelines
The process of building a machine learning (ML) model is iterative until you find the candidate model that is performing well and is ready to be deployed. As data scientists iterate through that process, they need a reliable method to easily track experiments to understand how each model version was built and how it performed.
Track your ML experiments end to end with Data Version Control and Amazon SageMaker Experiments
Data scientists often work towards understanding the effects of various data preprocessing and feature engineering strategies in combination with different model architectures and hyperparameters. Doing so requires you to cover large parameter spaces iteratively, and it can be overwhelming to keep track of previously run configurations and results while keeping experiments reproducible.
publications
Multi-hop MAC Implementations for Affordable SDR Hardware
J.C. O'Sullivan, P. Di Francesco, U.K. Anyanwu, L.A. DaSilva, and A.B. MacKenzie. 2011. "Multi-hop MAC Implementations for Affordable SDR Hardware." IEEE DySPAN, Aachen, Germany, pp. 632-636.
This paper presents the implementation and experimental evaluation of a MAC protocol designed to overcome some of the limitations of affordable and widely-deployed software defined radio hardware. We propose a modified Aloha-based MAC protocol with implicit acknowledgements to mitigate the impact of intra-flow collisions in multi-hop wireless communications. We experimentally observe a significant improvement in throughput and delay through simple modifications to the MAC protocol, tailoring it to the timing constraints of the USRP1.
Dynamic Spectrum Scheduling for Carrier Aggregation: A Game Theoretic Approach
Y. Xiao, Y. Chau, P. Di Francesco, and L.A. DaSilva. 2013. "Dynamic Spectrum Scheduling for Carrier Aggregation: A Game Theoretic Approach." IEEE ICC, Budapest, Hungary, pp. 2672-2676.
This paper addresses the problem of dynamic spectrum scheduling for carrier aggregation in heterogeneous wireless networks using a game-theoretic approach. We formulate the spectrum scheduling problem as a non-cooperative game among base stations and propose an algorithm to reach a Nash equilibrium for efficient carrier component assignment.
Customized Services over Virtual Wireless Networks: The Path towards Networks without Borders
L.A. DaSilva, J. Kibiłda, P. Di Francesco, T.K. Forde, and L.E. Doyle. 2013. "Customized Services over Virtual Wireless Networks: The Path towards Networks without Borders." Future Network and MobileSummit, Lisbon, Portugal.
Wireless networks of the future will be characterised by heterogeneity of spectrum usage regimes, of resource ownership models, and of radio access technologies. In these networks, resources will be orchestrated to create bespoke virtual networks designed to best meet the needs of specific services.
A Split Architecture for Random Access MAC for SDR Platforms
P. Di Francesco, S. McGettrick, U.K. Anyanwu, J.C. O'Sullivan, A.B. MacKenzie, and L.A. DaSilva. 2013. "A Split Architecture for Random Access MAC for SDR Platforms." CROWNCOM, Washington, DC.
Implementation of carrier-sensing-based medium access control (MAC) protocols on inexpensive reconfigurable radio platforms has proven challenging due to long and unpredictable delays associated with both signal processing on a general purpose processor (GPP) and the interface between the RF front-end and the GPP. This paper describes the development and implementation of a split-functionality architecture for a contention-based carrier-sensing MAC, in which some of the functions reside on an FPGA and others reside in the GPP. We experimentally test the performance of the resulting protocols in a multihop environment in terms of end-to-end throughput and required frame retransmissions.
Enhancing the Performance of Random Access MAC Protocols for Low-cost SDRs
A. Puschmann, P. Di Francesco, M.A. Kalil, L.A. DaSilva, and A. Mitschele-Thiel. 2013. "Enhancing the Performance of Random Access MAC Protocols for Low-cost SDRs." ACM WiNTECH, Miami, FL.
Software Defined Radio (SDR) is a technology which facilitates experimentation and the practical realization of novel wireless communication systems. Especially low-cost SDRs, however, experience high communication delays due to the connection between the radio hardware and the host computer. This delay hinders the implementation of Medium Access Control (MAC) protocols. In Carrier Sense Multiple Access (CSMA) based protocols, especially the Clear Channel Assessment (CCA) as well as the subsequent channel access phase are subject to strict temporal constraints. In this paper, we present two strategies that address both issues and aim to enhance the performance and efficiency of CSMA protocols implemented on low-cost SDRs.
Mobile Network Sharing Between Operators: A Demand Trace-Driven Study
P. Di Francesco, F. Malandrino, and L.A. DaSilva. 2014. "Mobile Network Sharing Between Operators: A Demand Trace-Driven Study." ACM SIGCOMM CSWS, Chicago, IL, pp. 39-44.
We assess, through real-world demand and deployment traces, how sharing can improve the efficiency of present-day cellular networks, especially in rural areas. We examine the main challenges and opportunities of infrastructure sharing and evaluate the potential gains from different sharing configurations between mobile network operators.
A Split MAC Approach for SDR Platforms
P. Di Francesco, S. McGettrick, U.K. Anyanwu, A.B. MacKenzie, and L.A. DaSilva. 2015. "A Split MAC Approach for SDR Platforms." IEEE Transactions on Computers 64(4): 912-924. doi: 10.1109/TC.2014.2308197
Implementation of carrier sensing-based medium access control (MAC) protocols on inexpensive reconfigurable radio platforms has proven challenging due to long and unpredictable delays associated with both signal processing on a general purpose processor (GPP) and the interface between the radio frequency (RF) front end and the GPP. This paper describes the development and implementation of a split-functionality architecture for a contention-based carrier sensing MAC, in which some of the functions reside on a field-programmable gate array (FPGA) and others reside in the GPP. We provide an FPGA-based implementation of a carrier sensing block and develop two versions of a carrier sense multiple access (CSMA) MAC protocol based upon this block. We experimentally test the performance of the resulting protocols in a multihop environment in terms of end-to-end throughput and required frame retransmissions. We cross-validate these results with a network simulator with modules modified to reflect the mean and variance of delays measured in components of the real software-defined radio system.
Infrastructure and Spectrum Sharing Trade-offs in Mobile Networks
J. Kibiłda, P. Di Francesco, F. Malandrino, and L.A. DaSilva. 2015. "Infrastructure and Spectrum Sharing Trade-offs in Mobile Networks." IEEE DySPAN, Stockholm, Sweden.
This paper investigates the trade-offs between infrastructure sharing and spectrum sharing in mobile networks. Using real-world deployment and demand data, we analyze how different sharing configurations impact network capacity, cost efficiency, and competition among mobile network operators.
A Sharing and Competition Aware Framework for Cellular Network Evolution Planning
P. Di Francesco, F. Malandrino, T. Forde, and L.A. DaSilva. 2018. "A Sharing and Competition Aware Framework for Cellular Network Evolution Planning." IEEE Transactions on Cognitive Communications and Networking 1(4): 464 - 470. doi:10.1109/TCCN.2017.2663060
Mobile network operators are facing the difficult task of significantly increasing capacity to meet projected demand while keeping CAPEX and OPEX down. We argue that infrastructure sharing is a key consideration in operators’ planning of the evolution of their networks, and that such planning can be viewed as a stage in the cognitive cycle. In this paper, we present a framework to model this planning process while taking into account both the ability to share resources and the constraints imposed by competition regulation (the latter quantified using the Herfindahl index). Using real-world demand and deployment data, we find that the ability to share infrastructure essentially moves capacity from rural, sparsely populated areas to urban ones, with significant increases in resource efficiency.
Sensitivity Analysis on Service-Driven Network Planning
P. Di Francesco, F. Malandrino, and L.A. DaSilva. 2017. "Sensitivity Analysis on Service-Driven Network Planning." IEEE/ACM Transactions on Networking 25(3): 1417 - 1430. doi: 10.1109/TNET.2016.2633417
Service providers are expected to play an increasingly central role in the mobile market and their relationship with the traditional mobile network operators (MNOs) is starting to change. The dilemma faced by over-the-top service-providers (OTTs) is now whether to enter into a service level agreement with the MNOs or to invest in deploying their own network infrastructure to serve their demand. The purpose of this paper is to study the factors shaping the agreements between OTTs and MNOs and how these factors impact network planning decisions. Using our model in conjunction with real-world data, we find that service-driven networks are heavily influenced by regulatory decisions, and that cost structures and demand characteristics play non-marginal roles in the definition of service-driven networks.
Assembling and Using a Cellular Dataset for Mobile Network Analysis and Planning
P. Di Francesco, F. Malandrino, and L.A. DaSilva. 2018. "Assembling and Using a Cellular Dataset for Mobile Network Analysis and Planning." IEEE Transactions on Big Data 4(4): 614 - 620. doi:10.1109/TBDATA.2017.2734100
In a world of open data and large-scale measurements, it is often feasible to obtain a real-world trace to fit to one’s research problem. Feasible, however, does not imply simple. Taking next-generation cellular network planning as a case study, in this paper we describe a large-scale dataset, combining topology, traffic demand from call detail records, and demographic information throughout a whole country. We investigate how these aspects interact, revealing effects that are normally not captured by smaller-scale or synthetic datasets. In addition to making the resulting dataset available for download, we discuss how our experience can be generalized to other scenarios and case studies.
research
talks
MLflow on AWS and SageMaker Integration
Organized and hosted the first AWS Vienna Meetup at the AWS Vienna office, presenting on MLflow integration with Amazon SageMaker and open-source ML tooling on AWS. The event reached full capacity with 35 attendees.
MLOps — Automazione per il Machine Learning
Recorded a 40-minute podcast episode on MLOps with Alex Casalboni for the official AWS Italy Podcast, discussing automation strategies for machine learning workflows, SageMaker Pipelines, experiment tracking, and best practices for production ML systems.
MLOps on AWS — Italian Track
Delivered an MLOps session as part of the Italian-language track at AWS Innovate, covering end-to-end machine learning workflows on Amazon SageMaker. Also participated in an Ask The Expert panel and live Q&A session. Approximately 200 attendees.
AWS DeepRacer Workshop
Delivered two hands-on AWS DeepRacer workshop sessions (theory and practice) at AWS Summit Milan, covering reinforcement learning fundamentals and autonomous racing. Both sessions were at full capacity with 30+ attendees each. Delivered in Italian.
Foundation Model Hosting and RAG on AWS
Presented on foundation model hosting options and demonstrated a Retrieval-Augmented Generation (RAG) application at the Vienna Data Science Tools Meetup, with approximately 60 attendees from the local data science community.
Productionize ML workloads using Amazon SageMaker MLOps
This talk presented a comprehensive overview of MLOps on AWS, covering the journey from experimental notebooks to production-ready ML systems using Amazon SageMaker. Starting from the premise that ML code is only a small fraction of a real-world ML system, the session walked through an MLOps maturity framework across four phases — Initial, Repeatable, Reliable, and Scalable — mapping each to specific AWS services and capabilities. Topics included SageMaker Studio for experimentation, SageMaker Experiments for tracking, SageMaker Pipelines for workflow automation, Model Registry for versioning and promotion, SageMaker Projects for one-click CI/CD provisioning, shadow testing and deployment guardrails, Model Monitor for drift detection, and Model Cards and Dashboard for governance. The talk also covered team structures, multi-account strategies, and custom project templates for enterprise-scale MLOps.
LLMOps and FMOps — Evaluating and Operating Foundation Models
Presented LLMOps and FMOps concepts to approximately 100 attendees at an AWS builder-focused event, covering the operational lifecycle of large language models including fine-tuning, evaluation, deployment, and monitoring strategies.
Accelerate FM pre-training on Amazon SageMaker HyperPod
Co-presented with Ankit Anand and Matt Nightingale, this session explored the challenges of training foundation models at scale and how Amazon SageMaker HyperPod addresses them. The talk covered the generative AI landscape and the growing computational demands of FM development, from prompt engineering and RAG to full pre-training. We introduced SageMaker HyperPod as a resilient, performant, and customizable environment for large-scale distributed training — featuring self-healing clusters that automatically detect hardware failures, replace faulty instances, and resume training jobs from checkpoints, reducing training time by up to 20%. The session went under the hood of HyperPod, covering cluster architecture, instance groups, lifecycle scripts, Elastic Fabric Adapter (EFA) for high-speed inter-node communication, distributed training software stacks for both GPU and Trainium, and job scheduling with auto-healing. Customer stories from Stability AI, Perplexity AI, and Hugging Face illustrated real-world benefits.
FMOps/LLMOps — Managing the Generative AI Lifecycle on AWS
Presented a 30-minute session on Generative AI lifecycle tooling to an audience of approximately 200 attendees at a cloud computing conference. Covered the end-to-end lifecycle of foundation models, from experimentation and evaluation to deployment and monitoring, using AWS services. The session led to a follow-up conference with business decision makers.