👨🏻‍💻 Currently pursuing my Master’s in Artificial Intelligence at Utrecht University, I combine my software engineering background with cutting-edge AI research to tackle real-world challenges. For this reason, I am working with MLLMs to help streamline the building energy retrofit process, which can save energy and promote sustainable living.

🔬 As I advance through my second year, I’m particularly excited about preemptive detection and monitoring of dangerous AI capabilities that arise from misalignment, aiming to create solutions that make a meaningful impact on society. In particular, I am interested in understanding LLM reasoning mechanisms such as Chain-of-Thought monitoring or enhancing them in the latent space as in Chain of Continuous Thoughts.

📚 My journey from software engineering to AI has been driven by a profound belief in technology’s potential to transform lives. Through my research and development work, I strive to bridge the gap between complex technical solutions and human needs, ensuring that AI serves as a force for positive change.

Selected Experience

🤖 Challenges

I am working as independent researcher for the AI Safety Camp where I Co-authored a paper about Bayesian Reasoning Elicitation in SoTA LLMs using prompt engineering.

I have implemented an LLM Agent for predicting salary, rent and subscription trends using logistic regression and LLama3-70B via Nvidia’s API. I guided my team in delivering a working prototype within 24 hours, resulting in a top 3 project by popular vote.

I have experience contributing to Predictive troubleshooting for medical devices, an hackathon project at the Utrecht Medical Center, for preventing infusion pumps from breaking using NLP with Neural approaches. Feel free to look at my post about it.

I worked for Improving Water level prediction using Exogenous variables and Neural Networks as part of an AI4LS2, a large hackathon project, proposed by the University of Vienna.

📜 Reimplementing and Reproducing Papers

Natural Language Processing

Developed a variant of the paper Finding Deceptive Opinion Spam by Any Stretch of the Imagination for my Data Mining course project.

Reproduced the paper Glove and validated the learned embeddings through their similarity. Feel free to look at my specific portfolio entry.

In the project Probing BERT: Understanding Language Model Behavior, the counterfactual conditionals tests are dapted from the paper Counterfactual reasoning: Testing language models’ understanding of hypothetical scenarios. This work focused on evaluating how BERT-based models reason about hypothetical scenarios, contributing to a deeper understanding of language model behavior in processing counterfactual information.

In the project Exploring Parsing Algorithms: Viterbi and CKY I have implemented the CKY Algorithm from the paper The Viterbi Algorithm

Computer Vision

Developed a small-scale variant of the paper Two Stream Neural Networks for my project Action Recognition in videos.

Medical Deep Learning

Implemented a small-scale variant of the paper Deep brain state classification of MEG data for my project MEG Data Analysis using Deep Learning.

Computational Neuroscience

I have implemented the tutorials from the masterclass Numerical and Analytical methods for spatially-extended neurobiological networks taught by the PhD D.Avitabile at the Utrecht University.

👨🏻‍🔬 Professional Experience

I have professional experience working as Software Engineer for a Japanese international company and United Arab Emirates’ startup after graduating from my Bachelor. I mainly worked to solve AI challenges and developed components for IT, Fintech and Booking companies.

🤝 Groups and Activities

I have joined the AI Safety Group in the Netherlands and, after completing the course AI Safety Foundamentals (2023 and 2024), I’m interested in working on the Alignment problems, whether it is about goal misgeneralization, reward mispecification, interpretability or privacy issues. In the final project for the 2023 cohort of AISF I have successfully showed how a DQN agent learns to play a game and misgeneralizes by carrying a misaligned behaviour using proxy features to achieve its goal.

I have participated in the workshop The numerical brain: forward and inverse problems in neuroscience applications at the Masterdam Vrije University. (21-23 October 2024)

To learn more about Supercomputers and I have participated in the Supercomputing course where I have done practical exercises on the Dutch national supercomputer Snellius. (October 2024)

I have participated in the Buddy Programme proposed by the Graduate School of Natural Science because I believe it’s important to socialize, challenge our ideas, and promote diverse thinking.