Experiences
Involved in designing and deploying cloud services and applications. I was responsible for deploying cloud services that powered AI applications, enabling access via REST APIs. Additionally, I helped in designing and developing serverless cloud-native applications.
Involved in designing and developing architecture for rules-based systems and distributed ETL solutions in anti-fraud and anti-money-laundering fields. Working on the development of business assets involving NLP techniques for process automation through the creation of internal libraries and REST API service and deploying them in multiple environments.
Research and development of multiple Deep learning models to solve Object Detection and Instance Segmentation tasks to identify, provided a set of pictures from a claim, a solid evaluation of the repair costs. Contributed to the design of the architecture for the deployment of the Deep learning solution.
Joined Prometeia as a Data Engineer in the Data Science business line. Involved in using programming and analytical techniques for the extraction, transformation and analysis of large datasets on distributed architectures. Development of ETL solutions for Big Data analysis, libraries, using mainly scala, python and spark.
Education
Industrial thesis in which deep learning approaches are used to recognize the section of the image containing the car, the identification of the visible car components, and the severity of any visible damage in order to quantify repair costs. The approach uses multiple Deep Learning and Computer Vision models to solve Object Detection and Instance Segmentation tasks and business domain knowledge to aggregate the models' results in meaningful output.
The approach proposed in this thesis aims to extract information from the graph structure of the pages of a website, using distributional semantic algorithms on random walks generated from each node, and combining this information with the textual information present in the html code in order to learn vector representations of web pages. The validation of this approach is carried out through a clustering task of the obtained vectors, producing encouraging results.