In my time at Ordbogen, I investigated the effect of low-bit weight quantization on a grammar error classification model (GECToR). I designed and developed a novel CUDA kernel that utilized the multiplication-free nature of ternary quantization. I also trained variations of this model with the ultra-low bit weight quantization applied, incorporating knowledge distillation and varius PTQ methods. I designed experiments and created a test bench to evaluate numerous relevant performance factors, including performance on downstream tasks and memory overhead. This work can be tracked through my blogs and my repository.

YBCO as a superconducting material is historically unreliable, and has been known to degrade as a result of various factors like humidity, temperature, solvent, and bonding material. I designed and built a system that autonomously measures the residual (room temperature) resistivity of the superconducting device, and tracks the effect of such factors on the residual resistivity. The goal of the project is to construct a model that will be able to use time course residual resistivity to predict whether or not a sample has degraded and the superconducting device would be non-functional. The benefit of such an approach is to avoid an expensive cool-down of the devices to test viability.

Superconducting materials are very delicate, and therefore the dust particulate in the air must be tracked. I designed an Arduino-based system of measuring particulate levels, as well as temperature and humidity, that uploaded the readings to a mySQL database through Arduino's WiFi capabalities. In the final product, I designed and milled a PCB for clean and clear connections between the devices. I also rewrote the initial device library to support I2C connection with multiple devices. 

In this project, we fully designed and ran an experiment at a music festival in July of 2023. We set out to identify how one's relationship with the environment (measured by a survey) was reflected in their art. We brought in materials for drawing, painting, and collaging, and we prompted people to create in such a unique environment. Ultimately, we have plans to display the art, extract the motifs both by hand and using computer vision, analyze the data that we collected, and release a paper. 

We also designed and deployed 3D modelled and AI generated representations of wildlife local to the Seattle area. This culminated in a poster and video.

Building on the work of the MIT course Biodiversity and Cities, another undergraduate and I researched the role that the BioBlitz (rapid ecological survey) has played in urban environments. We developed case studies with applications of ecological data in the spheres of urban planning, economic development, community outreach, and environmental protection. We also proposed a step-by-step guideline on how to best implement the BioBlitz within the city of Quibdó, Colombia. This project culminated in a paper and poster. 

I was involved with several projects pertaining to the study of torpor in mice. Under the supervision of Julian Roessler, I studied the metabolic profile of mice at different stages of entering torpor, conducting several time series assays investigating the role of fatty acid oxidation once body temperature drops below a certain level. We isolated this pathway in induced and natural torpor. I also began research on isolating the torpor neuronal pathways from the brain to the fat pads, culminating in a presentation. Finally, I built a database and pipeline for uploading and storing continuous temperature data from the mouse housing, cleaning the data and searching for statistical trends as a result of our experiments.