Where research meets its next generation
On 24 and 25 September, around 50 students of the Physics of Data master’s programme at the University of Padua gathered at the Asiago Astrophysical Observatory for the Physics of Data Workshop 2026, organised by the Department of Physics and Astronomy “Galileo Galilei”. Alongside faculty, research leads and speakers from academia and industry, Omnys took part twice over: as sponsor of the workshop’s closing dinner, and on stage, where ML Engineer Lorenzo Barbiero, himself a graduate of the programme, presented his work at Omnys on retrieval for private company documents.
From the lecture hall to the stage
Lorenzo completed his M.Sc. in Physics of Data at the University of Padua. His thesis, developed with Omnys, became a paper on synthetic data generation for domain-specific information retrieval, which received the Best Student Paper Award at IntelliSys 2025 in Amsterdam.
Today he builds retrieval and document-understanding systems at Omnys, running on client data in production. Returning to his own degree’s workshop, this time as a speaker, closed a loop: the methods these students study are the ones Omnys puts to work every day.
What the talk covered
The talk, Dense Retrieval Beyond the Benchmark, was built around one question, asked in Italian: how many days of leave do I have? The answer sat in an English HR policy that shared no words with the question, so traditional search methods found nothing. From that starting point, Lorenzo guided the students through real use cases: adapting AI search models to private documents by generating the missing training data with large language models, searching scanned pages directly from the image, and letting AI agents choose the search method for each question. In the published study, the fine-tuned open model beat a leading commercial service by 4.6 points in answer accuracy, with retrieval running on local hardware.
Research that reaches production
At Omnys, research is part of how products are built. The team publishes when results warrant it, and the same methods move into client systems: synthetic training data where no labelled examples exist, visual search where documents survive only as scans, and compact models that run on standard CPUs.
“Supporting the Physics of Data Workshop is an investment in the people who will build the next generation of AI systems. Students trained to model, measure and question their assumptions are exactly who we want working on our clients’ hardest problems.”
Davide Pozza, CTO, Omnys
Opportunities for students and graduates
After the talk, several students followed up with questions and to ask about internships. Omnys offers theses and internships on problems like those presented in Asiago, taken from the problem definition to a model running in production, with publication as a goal whenever the results warrant it.
“A few years ago I was sitting where these students sat. Coming back to show them where the same questions lead, in real systems used by real companies, meant a great deal to me.”
Lorenzo Barbiero, ML Engineer, Omnys
Students and graduates interested in a thesis or internship at Omnys can get in touch at careers centre