This project studies the "Seats in Room" problem as a declarative optimization task, where students must be partitioned into groups and assigned to seats under social, spatial, and relaxable constraints. I modeled the problem with Answer Set Programming using Clingo and Constraint Programming using MiniZinc/Gecode, precomputing Euclidean seat distances to avoid non-linear solver overhead during search. The report compares optimization hierarchies, search heuristics, timeout behavior, and solution quality, showing how conflict-driven ASP solving scales better on tightly constrained logical instances while lightweight CP heuristics remain useful as a reproducible baseline.

This project investigates few-shot semantic segmentation for industrial visual inspection, focusing on semiconductor wafer defect images where labeled masks are scarce and domain expertise is costly. The work combines a technical literature review of conditional networks, prototype-based methods, latent-space optimization, contrastive learning, and vision foundation models with the implementation of a generalist Transformer-based segmentation pipeline. I experimented with SegGPT through Hugging Face, using prompt images and masks to perform one-shot defect segmentation, and compared the approach with SAM2 prompting to analyze segmentation quality, prompt sensitivity, inference cost, and deployment trade-offs for AI-assisted quality control.

This survey analyzes the rapid development of Vision-Language Models after CLIP, focusing on how contrastive pretraining, multimodal representation learning, captioning objectives, and generative vision-language architectures changed the design space of computer vision systems. I compare representative models such as ALIGN, CoCa, FLAVA, LiT, and SLIP, emphasizing training objectives, data requirements, evaluation protocols, and the limitations of benchmark-driven progress. The goal is to map the technical trade-offs between transferability, semantic alignment, compositional reasoning, and deployment complexity in modern multimodal AI.

This data science project studies the role of nuclear energy in electricity production using public energy datasets from Ember, the Energy Institute Statistical Review of World Energy, and ElectricityMap. I built a comparative analysis pipeline to explore production trends, regional differences, and the relationship between nuclear generation, decarbonization, and energy-system stability. The project combines data cleaning, exploratory analysis, and visual communication to make energy-policy trade-offs inspectable from reproducible data sources.

During my Erasmus period in Spain, I developed a research thesis on predicting reactivity pulse values for a TRIGA nuclear reactor with neural-network models. The work frames reactor pulse estimation as a supervised learning problem in nuclear engineering, using deep learning to approximate safety-relevant physical behavior from experimental or simulated measurements. The thesis connects model design, training evaluation, and domain constraints with the practical objective of supporting reactor analysis workflows.

This project analyzes the psychological effects of the COVID-19 pandemic during a specific lockdown period through a statistical data-science workflow. The analysis covers data preparation, exploratory statistics, visualization, and interpretation of behavioral and stress-related variables. The project was evaluated with the highest grade, 30 cum laude.

In a university machine-learning competition, my team addressed income prediction as a supervised classification task. We compared feature preprocessing strategies and predictive models, optimizing accuracy and generalization under the constraints of the course challenge. The final system achieved 0.87 accuracy and ranked first in the competition.

I also developed a Kaggle notebook for breast-cancer classification, applying supervised learning to medical diagnosis data. The notebook focuses on executable preprocessing, model training, evaluation metrics, and reproducible experimentation for binary clinical classification.

In an autonomous robotics course during Erasmus, my team developed "Development of a Robot for Hazardous Environment Tasks" as the final project. The work included Python implementations of Kalman filters and particle filters for state estimation, together with proportional-integral-derivative controllers for robot motion control. The project connected probabilistic robotics, control systems, and practical navigation requirements in hazardous-environment scenarios.

Papel is an AI-assisted social network concept for academic reading and researcher interaction. The project explores on-device summarization, quiz generation, and privacy-preserving paper discovery, with the product architecture centered on local inference rather than server-side processing of reading behavior. It combines interface design with applied AI workflows for scientific literature navigation.

A personal space for technical notes, research reflections, and essays at the intersection of artificial intelligence, technology, art, and society.

My connection with technology began in 2013 when I started to understand how an android operating system works, therefore I started to create my own custom roms using as base the source codes of some popular aftermarket firmwares. In that time I met many people who contributed to the growth of my skills, @gokulbalram was my first guru and he deserves a special thanks. Like everything on this earth, this chapter of my life also came to an end, in 2017, after completing my latest job: an AOSP Porting for the ASUS Zenfone 2, powered by an Intel CPU. With the beginning of my university career, I deepened data science, generative and crypto art, blockchain and ethics related to technology. I like to work in dynamic, social environments where my creativity is encouraged. I believe that great goals can be achieved by reaching out to others, sharing your vision and valuing their input. The fundamental question I ask myself is: can an artificial intelligence in a near future think and act like a human? Make rational choices? Feels emotions?
being updated ...