Clayton Lemons

Used CUDA to accelerate my genetic algorithm for balanced train/test splits in NER datasets with multiple entity types. The fitness function builds histograms of entity counts, computes the Jensen-Shannon divergence between the training and test sets, and averages similarities across classes. The GA seeks the highest-scoring split. The fitness function is implemented as a CUDA kernel and packaged as a Python module with a C/C++ extension, compatible with PyGAD.

To beat a strong NumPy/SciPy baseline, mapped one CUDA block per candidate solution, parallelizing solutions across SMs. Batches from PyGAD are written once to GPU memory and launched as one block per solution. Wrote class-counts once to global memory and kept them resident across all solution evaluations.

Histograms live in shared memory with modest bin caps to stay within the memory budget. Threads compute JS terms in parallel from histograms. A custom per-class tree reduction sums those terms, and thread 0 computes the final fitness. Only solution permutations and results cross the PCIe boundary.

Performance landed around 110x speedup on realistic workloads (500 solutions, 100k samples) versus the optimized CPU path, with smaller batches still delivering 20–40x gains.

For my capstone project in a Scala functional programming course, I developed a Scala/Spark project to extract and process global temperature data. The goal was to create interactive, zoomable visualizations showcasing temperature patterns and deviations. This involved leveraging Scala for the functional programming aspects and Spark for handling large datasets efficiently, enabling dynamic and informative visualizations of climate trends worldwide.

I developed a Python/C++ client-server system for remote debugging and introspection of NI-DCPower API sessions across separate processes. The system features an optional server launched by the primary NI-DCPower session and a client that mirrors the session's API, essentially acting as a proxy. This system serves as the underpinning of the NI-DCPower Soft Front Panel Debug feature.

I implemented the project by creating a DLL that initiates an Apache Thrift server for command relay between remote and driver sessions over localhost, enabling seamless integration of debugging tools. Thrift's flexibility was key for future extensions to support connections from non-local processes.

I collaborated closely with LabVIEW and C# developers, who consumed the client API in order to implement the NI-DCPower Soft Front Panel Debug feature.

I created a Sublime Text plugin to integrate Perforce VCS features directly within Sublime Text while at National Instruments, aimed at boosting team workflow efficiency. This involved researching both Perforce and Sublime Text APIs and crafting features to streamline common developer tasks. This effort enhanced the development experience for my team members, who preferred using Sublime Text for their coding activities. I also worked with Sublime Text's Package Control community to get the plugin added.

I developed a prototype website for GMF Aquatics to showcase and sell their aquarium products, utilizing Amazon's merchant API. This pro bono project was intended to assist a friend in marketing their products more effectively. The website was engineered using CakePHP on the back end for scaffolding and rapid development. For the front end, I employed a combination of Bootstrap for a responsive design, Backbone.js for a structured web application, and Mustache for template rendering. This blend of technologies resulted in a simple but effective platform for highlighting and exploring aquarium products.