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, 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
-
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
-
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.
Physics Simulation → Data Generation → ML Training → Optimized C++ Inference → Scientific Visualization
Python · PyTorch · Qiskit · ONNX · C++ · Eigen · CMake · Gnuplot · LaTeX · Linux
GitHub: ksalamone59