Claudio Bonesana

Advanced Statistics and Data Mining Summer School

Big Data

January 1, 2016 – January 1, 2016

Scuola Universitaria Professionale della Svizzera Italiana (SUPSI)

Master of Science in Engineering, Intelligent System

January 1, 2013 – January 1, 2016

Scuola Universitaria Professionale della Svizzera Italiana (SUPSI)

Bachelor's degree, Ingegneria informatica

January 1, 2010 – January 1, 2013

Scuola Professionale Artigianale Industriale (SPAI)

Attestato Federale di Capacità, Informatica

January 1, 2009 – January 1, 2010

Liceo Cantonale Lugano 2

Maturià Liceale, Scientifico

January 1, 2005 – January 1, 2009

IDSIA USI-SUPSI

Lead Software Engineer: AI in Finance

November 1, 2024 – Present

IDSIA USI-SUPSI

Researcher in Software Engineering for Data Mining and Artificial Intelligence

March 1, 2016 – Present

IDSIA USI-SUPSI

MSc Students

September 1, 2013 – February 1, 2016

Flotta (formerly Ferdelance) - A Federated Learning framework for researchers

March 1, 2022 – September 1, 2024

Ferdelance is a federated learning based framework intended to be used both as a workbench to develop new distributed algorithm within a federated environment, and perform distributed statistical analysis on private data. Federated learning is a machine learning approach that allows for training models across decentralized devices or servers while keeping the data localized. Instead of collecting data from various sources and centralizing it in one location for training, federated learning enables model training directly on the devices where the data resides. This framework allows researchers to design, test, and train new federated learning algorithms in a simulated environment or in a real use case without rewrite the algorithm for a distributed setting.

Optimizing Job Ad Engagement: Leveraging Machine Learning for Enhanced Click-through Rates

January 1, 2022 – June 1, 2022

The industrial partner of this project was a leading job search engine aimed at optimizing ad allocation for enhanced job seeker engagement. The engine operates on a pay-per-click model, rewarding partner companies for each job ad click. In response to industry shifts towards pay-per-application structures, our goal was to refine ad allocation strategies to maximize application generation. My role involved developing a machine learning algorithm to ranking the available advertisements based on their click probability. Leveraging historical user data, the model was trained to assess the probability of a click on an offer. To engineer the features, several aspects were considered. First of all, geolocation data was utilized to understand the relation between a job offers and the different job markets around the world. These information were collected using geographic and statistical information regarding human and economic indicators across the globe. From the job offers, mixing NLP techniques and knowledge by the partner's experts, I extracted a number of features that indicates the most appealing part of the job ads. Metrics encompassing parnter's gains and costs, views, and clicks were also integrated for comprehensive analysis. Our prototype underwent rigorous testing against diverse portal behaviors, including instances of bot-driven clicks, capable of bring more gains to our partner but wasting clients money, and recruiter engagement variations, with less clicks overall but an higher engagement by their client. Results demonstrated that with just 10% of ads, up to 80% of clicks could be achieved, showcasing the algorithm's efficacy in selecting and optimizing the available job offers, creating personalized packages for the portals.

New Techno War

May 1, 2020 – June 1, 2021

In this project I was in charge of the development of the computer version of a tabletop strategic wargame made by Helvetia Games. The project focused on the implementation of a framework containing a simulator allowing the design, training, and evaluation of machine learning agents, and a web interface. I architected and implemented the simulation engine using Python, allowing interactive gameplay through a REST API or a simple Notebook. The user interface has been implemented with React, a framework that I studied just for the project, allowing an intuitive gameplay experiences for users, helping with the visualization and evaluation of agents in different scenarios. I implemented an interface for design new agents into the framework, enabling continuous enhancements to the pool of agents. This interface enable the crafting of sophisticated agents capable of strategic decision-making within the wargame environment. Starting from simple agents, based on greedy algorithms, I wrote more advanced agents that consider different aspects and understand rules of the game employing machine learning classification and prediction techniques, allowing more adaptive and predictive behaviors. I engineered a reinforcement learning infrastructure for agent training through a tourney-like environment, where historical data and self-learning mechanisms continously refine and optimize agent based on their performance. The selection of best agents was based on the ELO ranking. Quality of trained agents where good, although some limitations by the game's rules and scenarios levelled the overall performance. At the same time, I mentored a student throughout her semester project, where she designed and implementated diverse agents to enrich the wargame experience. I also supervised her master thesis focusing on Explainable AI (XAI), specifically exploring the rationale behind agents' decision-making processes and describing the underlying factors influencing action selection.

Fintech projects

November 1, 2018 – Present

I've accumulated experience within the fintech sector, particularly in collaboration with leading financial institution like UBS, on projects geared towards enhancing customer engagement and optimizing revenue generation through the integration of advanced technologies and data analytics. My contributions involved working on many initiatives aimed at developing sophisticated systems to improve customer interaction and promote proactive engagement strategies, and increase sales. In 2018, with the first iteration of this collaboration, I was selected with a small group of IDSIA researchers in proving the institute skills in multiple challenging projects with the partner. We achieved better performance in all projects respect to the existing solutions. In 2020 the collaboration has been renewed and I worked deeply tight with our partner's Natural Language Processing experts. Tasked in finding solution to unexplored areas of customer interaction, we build prototypes and experiments using state-of-the-art models. Since 2024, I took the Lead Engineer position fostering multiple projects each year with focus both on real case situation, experimentation of state-of-the-art models and algorithm, paper writing, and team management using cutting edge technologies based on Large Language Models, Agents and Agentic Frameworks, including safeguarding guardrails, and Reinforcement Learning. While my background underscores the convergence of fintech innovation and strategic business solutions, the confidential nature of my engagements prohibits me from disclosing specific project details.

Creation of a Recommender System for a regional Newspaper

April 1, 2018 – April 1, 2020

A regional newspaper in the southern Switzerland initiated a project with the goal of enhancing user traffic to their web platform and applications, while simultaneously boosting user engagement and the consumption of articles. The project culminated in the development of a recommender system designed to prioritize news items deemed relevant to the viewer, resulting in increased article views and overall web traffic. My role in the project encompassed three primary analyses. Firstly, I conducted feature extraction utilizing Natural Language Processing techniques applied to articles. Secondly, I performed user analysis and profiling within the platform, identifying distinct user behaviors and creating corresponding clusters. Lastly, I engineered a real-time recommender system capable of presenting users with a curated list of compelling articles. Following development, the prototype underwent rigorous testing in production using AB-test procedures to assess its impact, engagement levels, and effects on real users. Encouragingly, the results demonstrated an average 91% increase in click-through-rate compared to commercial state-of-the-art solutions.

Innovative SPC tools for a predictive TQM approach towards factory digitalization (TALOS)

March 1, 2017 – February 1, 2020

The TALOS Alarm System is a project in collaboration with GF Machining Solutions that aims to build a tool that introduces a preventive approach to the quality control of the assembly process of EDM machines. The tool will timely warn about possible risks of out of control parameters and it will empower the customization of the tests set of individual machines so to make easier the product customization required by the GF Machining Solutions's customers. In this project, my contribution was both on research and development parts. On the research side, I analyzed the assembly process, isolated the critical variables and norms to control, clustered them to recognize sets of correlated variables, and used this knowledge to design the structure of a Bayesian Network. This network it is used to predict possible deviations on the assembly process caused by the current state of the machine. This ensures an enhanced quality of the final product and reduce the risk of failure in the final quality tests and overall assembly cost. As part of the project, a prototype has been developed. The prototype consist in the development of a service that wraps the Bayesian network enabling at the same time its usage on the workshop.

CreMA - Credal Model Algorithms

March 1, 2017 – Present

In this project I did a complete rework of the structure in order to make it more resilient and with a more modern design based on interfaces. I also implemented some features like algorithms for distribution and networks sampling, imprecise and precise inference of Bayesian networks through message passing.

VIOLA - VIrtual Operators with self-Learning Ability

March 1, 2016 – March 1, 2018

VIOLA is a Artificial Intelligence based software that can operate in complete autonomy an electro-erosion Drill Machine. We developed this software during a KTI founded project in collaboration with the Research & Development department of +GF+ AgieChermilles and the Inspire group from ETH Zürich. The idea behind this project was to replicate the tests done by an human operator using the Drill Machine. These tests have the goal of find the optimal parameters for a new combination of alloys. VIOLA had the capabilities of find the optimal parameters given a database of test results or, if the combination of the materials wasn't already known, perform tests with a real machine and real materials to find the optimal parameters in an optimized way, by minimizing both the number of test to perform and the quantity of material consumed. In this project I had the task to build AMI: the Algorithm-Machine Interface. This was a communication interface between VIOLA (the learning algorithm) and the remote controlled Drill Machine. It also had the task to acquire the processing data before, during, and after the tests and make these data available to VIOLA. In other words, if VIOLA simulates the brain of an operator, AMI is its hand. I built this software using C# and some proprietary industry standard connection languages used by the machine (such as DNC for the action commands, Mt-Log and AC-Connect for reading status and measurements). This gave the possibility of use the existing features of the Drill Machines, such as, but not limited to, automatic electrode changing, manual electrode changing detection, sensors measurements. AMI was also capable of controlling multiple machines, and performing simultaneously different tests, thanks to a distributed system based on a client (the VIOLA software) and server (the machines) networks architecture.

GoEco

February 1, 2016 – November 1, 2019

The GoEco! project takes advantage of the wide availability of smartphones in order to overcome the traditional awareness-raising approach used to foster sustainable mobility and exploit eco-feedback, social norms and peer pressure elements in an ICT-based motivation system. In particular, it uses a smartphone app to analyze how we can encourage people to engage in more sustainable mobility lifestyles. Results have obtained by one month of large scale testing within the GoEco! living lab performed in Switzerland, allowing the collection of baseline mobility data for the sample of voluntary participants to the GoEco! living lab. Then, volunteers have been encouraged to use a more green approach to their mobility during two different stages. In this project I was in charge of building and maintaining the software API interface for building and training a classification model used to identify a mean of transport based on the mobility data collected for each user. This software was part of a distributed application used to operate the GoEco! application and produce in real time alternative green trips to the volunteers. Additionally, I was also in charge of the production and dissemination of personalized reports used to summarize the performance of each volunteer, and the aggregated analysis of the changes in the mobility of the sample participants.

Master Thesis - A News Clustering Software based on Data Mining Techniques

June 1, 2014 – January 1, 2016

This was a joint project involving as a research partner, the Dalle Molle Institute for Artificial Intelligence (IDSIA), and Newscron, a Swiss start-up working on the field of web news analytics, as industrial partner. The goal of the partnership was to apply data-mining techniques developed by IDSIA to the news aggregation system used by Newscron. This task has been formalized as an online supervised clustering problem. After a preliminary analysis of the state of the art, existing approaches based on the K-Means algorithm and Support Vector Machines have been considered. A new algorithm based on probabilistic graphical models and a Bayesian voting system has been indeed developed in order to cope with the dynamic stream of online news processed by the Newscron framework. The new method competes with the state of the art and significantly improves (+15%) the quality of the aggregation performed. Notably, the new method is also faster (-30%) and it allows to process more news articles in the same time. Because of these improved performances, the algorithm has been used to implement a new aggregation engine, already adopted by the Newscron servers.

Bachelor Thesis - STOUCH - Real Time Object Identification - Computer Vision

May 1, 2013 – September 1, 2013

Questo progetto prevede la progettazione e la realizzazione di un prototipo per sistema autonomo che permetta ad utenti, identificati da un braccialetto dotato di tag RFID, di scattare fotografie in luoghi strategici dal profilo turistico, a pubblicarle e condividerle attraverso i nuovi canali di comunicazione interattivi quali smartphone, tablet, blog e social networks. Dal profilo tecnico, il sistema realizzato è incentrato su un Raspberry Pi, con core ARM11, collegato con periferiche interattive e multimediali. Il sistema software scelto è basato su un sistema Linux derivato da Debian con librerie Qt e OpenCV per la manipolazione delle immagini. Valore aggiunto è l’utilizzo di un sistema di cross-compilazione e debugging remoto dell’applicativo software. Il progetto ha portato alla realizzazione di un prototipo funzionante che è stato testato sul campo.

Convolutional Neural Networks

Coursera

June 24, 2026 – Present

Sequence Models

Coursera

June 24, 2026 – Present

Deep Learning Specialization

Coursera

June 24, 2026 – Present

Natural Language Processing Specialization

Coursera

June 24, 2026 – Present

Build Basic Generative Adversarial Networks (GANs)

Coursera

June 24, 2026 – Present

Generative Adversarial Networks (GANs) Specialization

Coursera

June 24, 2026 – Present

Neural Networks and Deep Learning

Coursera

June 24, 2026 – Present

Natural Language Processing with Sequence Models

Coursera

June 24, 2026 – Present

Natural Language Processing with Attention Models

Coursera

June 24, 2026 – Present

Apply Generative Adversarial Networks (GANs)

Coursera

June 24, 2026 – Present

Machine Learning Engineering for Production (MLOps) Specialization

Coursera

June 24, 2026 – Present

Natural Language Processing with Classification and Vector Spaces

Coursera

June 24, 2026 – Present

Natural Language Processing with Probabilistic Models

Coursera

June 24, 2026 – Present

Build Better Generative Adversarial Networks (GANs)

Coursera

June 24, 2026 – Present

Machine Learning Modeling Pipelines in Production

Coursera

June 24, 2026 – Present

Introduction to Machine Learning in Production

Coursera

June 24, 2026 – Present

Machine Learning Data Lifecycle in Production

Coursera

June 24, 2026 – Present

Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization

Coursera

June 24, 2026 – Present

Structuring Machine Learning Projects

Coursera

June 24, 2026 – Present

Deploying Machine Learning Models in Production

Coursera

June 24, 2026 – Present