Alessandro Potenza
Computer Engineering student at Politecnico di Milano, building high-performance AI systems, GPU kernels, and open-source tools.
Featured Projects
AI/ML2026
agentrial
The pytest for AI agents. Run your agent 100 times, get confidence intervals instead of anecdotes. Published open-source framework with Wilson confidence intervals, step-level failure attribution via Fisher exact test, and real cost tracking across 45+ models.
Systems2026
Flash-Reasoning
Tree-Aware KV-Cache Attention for Reasoning LLMs. Custom Fused GQA Triton kernels exploit physical prefix sharing to exceed HBM bandwidth limits, enabling efficient inference for System 2 Reasoning models like DeepSeek-R1.
Systems2026
Flash-SAE
High-Performance Triton Kernels for Sparse Autoencoders. 13.6x speedup and 97% memory reduction via sparse kernel fusion. Drop-in PyTorch replacement with full autograd support for Mechanistic Interpretability research.
AI/ML2026
Verify-CBL
Neuro-Symbolic Formal Verification Engine combining Z3 SMT solver with LLM-powered code translation. Mathematically proves behavioral equivalence between legacy and modernized code, detecting 'penny drift' that testing misses.
All Projects
AI/ML2026
agentrial
The pytest for AI agents. Run your agent 100 times, get confidence intervals instead of anecdotes. Published open-source framework with Wilson confidence intervals, step-level failure attribution via Fisher exact test, and real cost tracking across 45+ models.
Systems2026
Flash-Reasoning
Tree-Aware KV-Cache Attention for Reasoning LLMs. Custom Fused GQA Triton kernels exploit physical prefix sharing to exceed HBM bandwidth limits, enabling efficient inference for System 2 Reasoning models like DeepSeek-R1.
Systems2026
Flash-SAE
High-Performance Triton Kernels for Sparse Autoencoders. 13.6x speedup and 97% memory reduction via sparse kernel fusion. Drop-in PyTorch replacement with full autograd support for Mechanistic Interpretability research.
AI/ML2026
Verify-CBL
Neuro-Symbolic Formal Verification Engine combining Z3 SMT solver with LLM-powered code translation. Mathematically proves behavioral equivalence between legacy and modernized code, detecting 'penny drift' that testing misses.
Research2025
GPU Performance Analysis: Triton vs. CUDA
Research paper quantifying the performance gap between high-level GPU programming models (Triton) and hand-optimized CUDA kernels, with a focus on irregular workloads like finite state machines.
PythonC++CUDATriton+1
Research2024
SplatSLAM
Real-time 3D mapping and SLAM from monocular RGB video using 3D Gaussian Splatting. Photo-realistic dense reconstruction without depth sensors, built as a Nerfstudio extension.
Web2024
ConceptHub
AI-Powered full-stack learning platform integrating Google's Gemini API. Automates generation of book summaries and conceptual mind maps from text, with user authentication and persistent storage.
AI/ML2024
Music Genre Classification
End-to-end reproducible pipeline achieving SOTA 83.5% accuracy on GTZAN with a U-Net inspired model. Leak-free methodology with track-level splits, cross-validation, and transfer learning.
Web2023
Chessboard.js
Modern, dependency-free JavaScript library and NPM package for building interactive chess experiences. Rich API for programmatic control, drag-and-drop, animations, and legal move enforcement.
Skills
Programming
Python · C++ · Java · SQL · JavaScript · TypeScript · C
AI/ML
PyTorch · TensorFlow · Scikit-learn · LLMs (Gemini API) · NLP · Computer Vision · SLAM
Systems
Docker · Kubernetes · Linux · Git · GCP · AWS · CUDA · Triton
Databases
PostgreSQL · MongoDB · MySQL · Redis · VectorDB · Spark
Web
React · TypeScript · Node.js · REST APIs
Tools
VS Code · Jupyter · Vim
Experience
Education
Achievements
Merit-Based Scholarship
Dec 2025Politecnico di Milano
Awarded specifically for outstanding academic performance (GPA and credits) during the M.Sc. in Computer Science and Engineering.
Global Finalist - Tech4Good
2024Huawei Technologies
Selected for the elite 'Seeds For The Future' program. Led a team to design an AI computer vision prototype, winning the National Competition and advancing to the Global Finals in China.
3rd Place - AI Challenge
2025Polimi (AIRLab)
Ranked 3rd out of 193 teams (Top 1.5%) in the Artificial Neural Networks Challenge. Engineered a custom Vision Transformer ensemble for complex medical image classification.