Algorithm and computational geometry engineer focused on performance-critical systems, including interactive geometry engines and CUDA-accelerated simulations. My work emphasizes mathematical correctness, deterministic behavior, and scalable computation over framework-heavy application code.

• Geometry kernels for interactive systems
• Constraint-aware editing and topology preservation
• Graph-based reasoning for planning and layout problems
• Analytic geometric evaluation without approximation shortcuts

• CUDA kernel development and parallel evaluation pipelines
• GPU memory layout and throughput optimization
• Deterministic numerical simulations on the GPU

• Real-time tools where geometric validity is preserved continuously
• Freeform editing with snapping, constraints, and invariants
• Separation of interaction, evaluation, and computation layers

• Feature-engineered OCR systems combining classical machine learning and deep learning approaches
• Custom dataset generation pipelines and synthetic data tooling
• Spatial feature extraction and representation design for tree-based models
• OCR systems emphasizing interpretable feature design, dimensionality reduction, and computational efficiency

CUDA-accelerated physics engine for large-scale electric field evaluation and field-line integration. Focused on numerical stability, parallel execution, and scalable vector-field computation independent of rendering concerns.

Interactive floorplan canvases and heuristic layout solvers combining computational geometry, graph algorithms, and constraint reasoning for automated layout generation and validation.

Freeform Bézier editors, planning engines, and analytic geometry utilities built on domain-specific geometry kernels without relying on dense polygon discretization.

OCR project exploring how far engineered feature representations can push traditional machine learning. Reduced handwritten character samples from 441 raw pixels to a compact 64-feature representation and achieved over 99.5% accuracy on moderately noisy data while approaching deep residual network performance.

Custom WPF application for generating MNIST-style datasets directly from Google Fonts. Designed for deterministic dataset generation and PyTorch-compatible export workflows.