• Research focused on Natural Language Processing (NLP) and Information Retrieval (IR), with applications to large-scale multimodal and text-based retrieval systems.
• Developed and evaluated novel deep learning architectures for language understanding and retrieval efficiency.
• Collaborated with international research groups on applied machine learning and multimodal AI projects.

• Specialized in Big Data, Machine Learning, and Statistical Modeling.
• Completed thesis research on large-scale data analysis and predictive modeling.
• Engaged in interdisciplinary coursework bridging computer science, mathematics, and data-driven decision-making.

• Built strong foundations in Algorithms, Databases, and Web Development.
• Developed several software projects in JavaScript, Python, and Java, emphasizing scalable backend systems.
• Graduated with distinction, consistently achieving top marks in core computer science courses.

This paper proposes three pre-training objectives designed to mimic the downstream fine-tuning task of contextual Answer Sentence Selection, and shows that their pre- training approaches (applied to RoBERTa and ELECTRA) can improve baseline contextual AS2 accuracy by up to 8% on some datasets.

Notes:

• State-of-the-art results on TrecQA and WikiQA benchmarks.

This paper proposes three novel sentence-level transformer pre-training objectives that incorporate paragraph-level semantics within and across documents, to improve the performance of transformers for Answer Sentence Selection, and mitigate the requirement of large labeled datasets.

This paper shows that popular pre-trained transformers perform poorly when used for fine-tuning on multi-candidate inference tasks, and proposes a new pre-training objective that models the paragraph-level semantics across multiple input sentences.

Self-supervised pre-training tasks that align structurally with downstream applications, reducing the need for labeled data and achieving state-of-the-art results on various benchmark datasets are proposed.

It is proved that eliminating the MASK token and considering the whole output during the loss computation are essential choices to improve performance and it is shown that ELECTRA benefits heavily from a state-of-the-art hyper-parameters search.

A main problem with reproducing machine learning publications is the variance of metric implementations across papers. A lack of standardization leads to different behavior in mechanisms such as checkpointing, learning rate schedulers or early stopping, that will influence the reported results. For example, a complex metric such as Fréchet inception distance (FID) for synthetic image quality evaluation (Heusel et al., 2017) will differ based on the specific interpolation method used.

Notes:

• TorchMetrics packages is downloaded about 300K times a day (source PyPI)

Constrained Adversarial Networks are proposed, an extension of GANs in which the constraints are embedded into the model during training and which efficiently generate valid structures that are both high-quality and novel.

This chapter introduces the semantic loss, a training method that incorporates symbolic structural constraints into neural networks to ensure outputs respect underlying dependencies (like valid graph paths). We enhance it with entropy minimization to prefer simpler solutions and demonstrate its versatility by integrating it with both discriminative and generative models, enabling efficient generation of complex structured objects.