Applied Scientist

Applied science sits between pure research and pure engineering. An applied scientist needs to be rigorous enough to produce results that can be published and reproduced, and practical enough to build the system that produces those results. Most people are strong at one of these. My background required both - research at VIT, Auburn, and ASU produced four publications, and the same pipelines that generated those results were production-grade engineering artifacts.

The standard I hold myself to is: could someone else reproduce this result on their own data? That requires documentation of every preprocessing step, versioned datasets, fixed random seeds, and evaluation protocols that do not overfit to the specific test set you happen to have. This is not standard practice in most academic ML labs, but it is the standard that separates applied science from research that does not transfer.

At ASU, I currently apply this standard to geospatial computer vision for sustainability research - processing satellite imagery with spatial fidelity quality gates that make the pipeline self-validating. Every output is checked against PSNR, SSIM, and RMSE thresholds before downstream inference runs on it. That is what production-ready research infrastructure looks like.

My research has focused on ML applications in clinical settings - areas where precision is not negotiable and where a model that fails silently has direct patient impact. At VIT CHAIR over two research internships, I built deep learning pipelines for Parkinson's disease detection from clinical signals, diabetic retinopathy classification from retinal images, and automated sleep apnea classification using variational mode decomposition combined with statistical features extracted from polysomnography data.

Each of these problems required solving a different challenge. Parkinson's detection from signal data required careful preprocessing to remove artifacts before any model sees the data. Retinopathy classification required handling significant class imbalance in clinical datasets. Sleep apnea classification required feature engineering from raw physiological signals before any ML approach was applicable. The work produced four publications and established the pipelines that were cited in subsequent research.

At Auburn University, I built the experiment tracking infrastructure for the retinopathy work - making every architectural comparison between CNN variants reproducible and reportable with appropriate statistical rigor. That infrastructure is what turned exploratory research into publishable findings.