I am an AI Engineer working at the intersection of engineering and AI/ML. This profile collects my work in GPU computing, deep learning, and LLM/Vision-LLM benchmarking.

Can small Vision-LLMs tell when a document question can't be answered — by actually reading the document, not guessing? Built a synthetic corruption benchmark on DocVQA (entity/element/layout corruptions), evaluated 3 small Vision-LLMs (2B–4B), and tested 3 in-context mitigation strategies across single- and multi-page settings. Highlight: the smallest model (2B) beat both larger ones — scale wasn't the advantage here. Vision-LLMs Qwen2-VL Gemma-3 Phi-3.5-Vision spaCy LLM-as-a-judge

Benchmarked proprietary LLMs (GPT, Gemini, Codestral) against open-weight models (Qwen 4B/8B, Phi-4-mini-reasoning) on natural-language-to-SQL generation, graded across a taxonomy of 7 SQL difficulty patterns and 4 database schemas — including full reasoning-trace evaluation, not just final-answer accuracy. Text-to-SQL LLM Evaluation Prompt Engineering Reasoning Analysis

Malware vs. benign classification from Windows API-call sequences, comparing a Bag-of-Words + Random Forest baseline, an embedding-based FFNN, and an LSTM (PyTorch). Highlight: the frequency-only baseline outperformed both sequential/neural models on macro F1 — model complexity should match the actual signal in the data, not just what's fashionable. PyTorch Scikit-learn LSTM Embeddings Imbalanced Classification

Hands-on GPU performance work: NumPy → CuPy → custom CUDA kernels (Numba), then profiling and tuning with NVIDIA Nsight Systems — block-size sweeps, parallel reductions, and diagnosing synchronization anti-patterns directly from the profiler timeline. CUDA Numba CuPy Nsight Systems GPU Profiling

My Master's thesis: evaluating the new SysML v2 language for Model-Based Systems Engineering in aerospace, with a liquid hydrogen tank case study and Simulink interoperability — in collaboration with Leonardo S.p.A. SysML v2 MBSE Simulink Systems Engineering

AI/ML: PyTorch · Scikit-learn · Transformers (Qwen, Gemma, Phi) · Prompt Engineering GPU/Performance: CUDA · Numba · CuPy · NVIDIA Nsight Systems Engineering tools: MATLAB · SolidWorks · Siemens NX · SysML v2 · Simulink Languages: Python · SQL

⭐ Feel free to explore the repos above — each has its own README with methodology and key results.