Experimental Physics PhD student working on large-scale detector data analysis, machine learning systems, and computational physics. Focused on designing and deploying scientific and ML pipelines in high-performance research environments, with emphasis on quantum information and data-driven modeling of physical systems.
Current PhD research is proprietary — projects here represent independent work.
- High-performance computing
- Quantum and computational physics (variational algorithms, Hamiltonian simulation)
- Machine learning systems engineering (PyTorch → ONNX → C++ inference pipelines)
- Scientific computing infrastructure (reproducible visualization and analysis tooling)
- Developed a backend-agnostic C++ inference profiler (ONNX Runtime + LibTorch) with zero-allocation hot path; achieved 13.97M samples/sec and R² = 0.9999 cross-backend numerical agreement
- Built C++ ML inference system achieving 9.4M samples/sec (16× speedup) in batched vs per sample inference
- Quantified VQE error decomposition (discretization vs variational limits)
- Designed end-to-end scientific pipelines (simulation
$\rightarrow$ ML$\rightarrow$ deployment$\rightarrow$ visualization)
-
Built a backend-agnostic C++ framework for benchmarking and comparing ML inference engines in Python-free deployment environments. Supports ONNX Runtime and LibTorch through a common abstract interface, CPU and GPU inference modes, configurable threading, reusable tensor allocation, statistical benchmarking (Welford online variance), backend comparison, and automated batch-size sweep studies.
-
Measured up to 13.97M samples/sec using ONNX Runtime with four threads and demonstrated R² = 0.9999 agreement between ONNX Runtime and LibTorch backends.
-
Why this matters: Real-world ML deployment often requires choosing between inference runtimes with competing performance, portability, and maintenance tradeoffs. This project provides a reproducible framework for quantifying those tradeoffs, validating numerical consistency across backends, and identifying bottlenecks before deployment into performance-critical environments.
Focus: C++ systems engineering, ML deployment, performance optimization, benchmarking methodology, backend abstraction, GPU vs CPU inferencing
-
C++ inference engine with pre-allocated tensor reuse and singleton session management for zero-overhead-per-call ORT deployment. Includes statistically rigorous benchmarking via Welford online variance estimation — batched inference achieves 9.4M samples/s, ~16× over sequential baseline. Engineering patterns drawn from production physics reconstruction constraints.
-
Why this matters: Demonstrates the complete path from model development in Python to high-performance deployment in a production-style C++ environment, while quantifying the performance impact of batching, memory reuse, and inference-engine design choices.
Focus: ML deployment, C++ inference systems, performance benchmarking
-
Variational Quantum Eigensolver (VQE) study of the hydrogen atom ground state, directly benchmarked against a classical eigenvalue solver under identical finite-difference discretization.
-
Why this matters: Separates algorithmic limitations from numerical discretization effects, helping clarify where quantum resources provide meaningful improvements and where classical approximation error dominates observed performance.
Focus: quantum algorithms, Hamiltonian discretization, error decomposition, variational landscapes, scaling behavior
-
Developed a geographically aware linear programming model using SciPy's HiGHS solver to minimize transportation costs across 5 warehouses and 20 customer demand zones while satisfying warehouse capacity and customer demand constraints.
-
Modeled transportation costs as a function of Euclidean distance between warehouses and destinations, generating optimal shipment allocations and network visualizations using NumPy and Matplotlib.
-
Implemented resilience and sensitivity analysis by simulating individual warehouse closures and 50% capacity reductions, quantifying changes in total network cost and identifying the most critical warehouse facilities.
-
Built a reproducible project workflow with a Makefile, virtual-environment dependency management, automated plot generation, and deterministic simulation using NumPy's random number generator.
-
Why this matters: Demonstrates the application of mathematical optimization and computational modeling to a practical resource-allocation problem, including analysis of how an optimized network responds to infrastructure disruptions.
Focus: linear programming, operations research, optimization, sensitivity analysis, computational modeling, scientific visualization
-
Reproducible gnuplot + LaTeX system for consistent publication-quality scientific figures across projects.
-
Why this matters: Reproducible visualization infrastructure reduces manual figure generation, improves consistency across projects, and makes scientific results easier to verify, maintain, and communicate.
Focus: scientific visualization, automation, reproducibility
Results from the C++ ML Inference Profiler Engine showing throughput as a function of batch size for ONNX Runtime and LibTorch backends under single-threaded and four-threaded execution. ORT consistently outperforms LibTorch at the same thread count. Threaded variants have higher variance due to synchronization overhead.
This heatmap shows the output from characterizing VQE as a solution to the Hydrogen atom's ground state. Quantifies the minimum achievable error as a function of qubit count and maximum radius r in the Hamiltonian discretization.
Output from the PyTorch → ONNX → C++ inference pipeline showing: - Noisy input data to the C++ inference - The output C++ inference - The true function
All figures shown above were generated using the Scientific Plotting Infrastructure repository.
Network representing the optimized solution for the supply chain optimizer. The blue squares are the warehouses, orange points the customer destinations, and the size of the lines connecting them is a proxy for the amount of resources being transferred.
## System ViewPhysics Simulation → Data Generation → Optimization / ML → High-Performance Inference → Scientific Visualization
Python · NumPy · SciPy · Matplotlib · PyTorch · Qiskit · ONNX · C++ · Eigen · CMake · Gnuplot · LaTeX · Linux
GitHub: ksalamone59