MSc Computer Engineering @ Politecnico di Milano · AI Systems & GPU Programming

I build high-performance systems at the intersection of deep learning and GPU programming. My work spans custom CUDA/Triton kernels for LLM inference, formal verification with SMT solvers, and statistical frameworks for AI agent evaluation.

agentrial — The pytest for AI agents: run your agent 100 times, get confidence intervals instead of anecdotes. (PyPI · VS Code)

Flash-Reasoning — Tree-Aware KV-Cache Attention for Reasoning LLMs, with custom fused GQA Triton kernels exceeding HBM bandwidth limits.

Flash-SAE — High-performance Triton kernels for Sparse Autoencoders. 13.6x decoder speedup, 97% memory reduction.

Verify-CBL — Neuro-symbolic formal verification engine using Z3 SMT solver and LLM-powered code translation.

SplatSLAM — Real-time 3D SLAM from monocular RGB using 3D Gaussian Splatting. No depth sensors needed.

ConceptHub — AI-powered learning platform with auto-generated book summaries and mind maps.

Music Genre Classification — SOTA 83.5% accuracy on GTZAN with a U-Net inspired model and leak-free methodology.

Chessboard.js — Dependency-free JS library for interactive chess: drag-and-drop, animations, legal move enforcement.

• Merit-Based Scholarship, Politecnico di Milano
• Global Finalist, Huawei Seeds for the Future: led a team to the Global Finals in China
• Top 1.5% (3rd/193) in the Polimi AI Challenge with a custom Vision Transformer ensemble
• 110/110 cum Laude, Sapienza University of Rome, Honors Program (Top 1%)
• Technical book author for a Computer Science textbook (Neldiritto Editore)

Python · C++ · CUDA · Triton · PyTorch · Docker · TypeScript · React · PostgreSQL