I am a Data Scientist with a unique background in Cognitive Neuroscience. My approach to AI is deeply rooted in the study of complex systems: I don't just see rows of data; I see behavioral patterns.I treat data as a dynamic signal to be decoded and make it actionable.
Leveraging my background in Cognitive Neuroscience, I treat data as a dynamic signal with a rigorous experimental mindset:
- Behavioral profiling: Treating machine telemetry as a proxy for system "health behavior."
- Pattern recognition: Identifying hidden "rhythms" in noisy, high-volume datasets.
- Explainability (XAI): Making complex models transparent and actionable for human stakeholders.
- End-to-end ML pipelines From raw, noisy data preprocessing to modeling, tuning, explainability and interactive visualization
| Project | Challenge and data | Tech stack (algorithms and tools) | Impact and knowledge discovery |
|---|---|---|---|
| Predictive Maintenance for Urban Mobility | 7.5M Telemetry Logs. Transitioning gates from "fail-and-fix" to proactive maintenance. | ML: XGBoost, CatBoost, LightGBM, RF. DL/TS: LSTM, MiniRocket, sktime. XAI: SHAP. |
Built a Stakeholder Dashboard (Streamlit/Plotly). Isolated maintenance noise from operational stress for Root-Cause Analysis. |
| Normative Graph-RAG for Aerospace Compliance | Aerospace Standards (ECSS). Overcoming semantic ambiguity and deontic rigidity (SHALL/SHOULD) in complex regulatory frameworks. | Data Engine: NetworkX, ChromaDB. ML/LLM: LangChain, Hugging Face (Qwen2.5), Sentence-Transformers. Visuals: Streamlit, Pyvis. |
Developed a Hybrid RAG Pipeline merging semantic vector search with a Normative Document Graph. Built an interactive Sub-graph Explorer ensuring grounded retrieval and resolving topological normative cross-references. |
| EU Diagnostic Imaging Market Analysis | Eurostat Healthcare Data (CT, MRI, PET). Analyzing examination volumes across 8 European markets to guide contrast agent portfolio strategy. | Data Engine: Python (Pandas, SciPy). BI & AI: Power BI (DAX, Key Influencers). Forecasting: Exponential Smoothing. |
Delivered an interactive Executive Intercative Report. Applied Z-score anomaly detection to isolate COVID-19 reporting noise. Identified the Italian penetration gap and structural shifts toward MRI/PET diagnostics for strategic commercial planning. |
| Alzheimer’s Clinical Support System | Medical Diagnostics. Early detection of dementia using clinical and lifestyle datasets. Addressing class imbalance and demographic biases. | ML: LightGBM (Baseline vs Fin-etuned), CatBoost, XGBoost, Random Forest. XAI: SHAP (Waterfall Plots). Visuals: Streamlit, Plotly (Radar Charts). |
Developed a Diagnostic Support Dashboard. Prioritized Clinical Integrity by selecting baseline LightGBM over fine-tuned variant and reducing overfitting for superior generalizability. Visualized Patient Profile vs. Population Benchmarks to facilitate evidence-based clinical decisions. |
| Wearable stress and affect: interpretable physiological ML | WESAD multimodal recordings from 15 participants. Classifying baseline, stress, and amusement from wrist and chest signals while investigating individual variability and subjective experience. | Signal Processing: NumPy, Pandas, SciPy. ML: Logistic Regression, Random Forest, XGBoost, LightGBM, MLP; nested LOSO validation. Interpretability: Monte Carlo Shapley, beeswarm/waterfall plots, PDP, ALE. Statistics and visuals: Spearman, paired bootstrap, Matplotlib, Seaborn. |
Built a subject-independent benchmarking pipeline connecting physiological features, self-reports, and model reliability. Exploratory late fusion reached 73.3% balanced accuracy on the three-class task; wrist logistic regression reached 90.3% on stress/non-stress. Investigated feature redundancy, modality contributions, and prediction abstention, distinguishing laboratory performance from real-world validity. |
| Motor Imagery BCI: Decoding Neural Intentions | Brain-Computer Interface. Decoding motor imagery from EEG signals using the BCI Competition IV-2a dataset. | Neuroscience & Signal Processing: MNE-Python, ICA, CAR, band-pass filtering. ML: CSP, LDA, scikit-learn,Riemannian Geometry. Deep Learning: EEGNet, Braindecode, TensorFlow/Keras, PyTorch. Deployment: Streamlit, PyLSL. |
Developed an EEG-based BCI pipeline integrating signal preprocessing, artifact mitigation, classical machine learning, deep learning, and Riemannian geometry. Implemented binary and four-class motor imagery decoding, cross-subject and cross-session evaluation, and interactive Streamlit demos. Visualized **CSP spatial patterns and EEG topographies ** to investigate the physiological plausibility of learned features around sensorimotor regions (C3/C4). |
| Shelter Intelligence: Adoption Optimization | Social AI. Identifying demographic and intake factors influencing animal adoptability. | Clustering: K-Means,DBSCAN,Hierarchical . Classification and Regression: Random Forest, Decision Trees, KNN, SVM, LinearRegression. |
Analyzed factors impacting stay duration and adoption outcomes to optimize resource allocation in shelters. |
| NeuroPsy-Research Archive | Behavioral Data Modeling and Experimental Design. Mapping clinical traits to cognitive and perceptual performance and body perception. | PsyToolkit, SPSS, Excel. Paradigms: Posner Task, Memory Task, Shelf Task, ECR- 12, AST, STA-Y2,LCSQ, BDS |
Engineered automated scoring for clinical scales and biometrics. Validated "Healthy is Up" metaphor via spatial categorization and BMI correlations. Mapped attachment style-driven biases in attention (Posner), memory (Old-New), and social interpretation (AST) |
- Master in Big Data Analytics and AI for Society | University of Pisa
- MSc in Cognitive Neuroscience and Clinical Neuropsychology | University of Campania
"Treating data not just as numbers, but as the behavior of a complex system."