Machine Learning Engineer

At Omdena from July to October 2024, I led ML engineering for a health and longevity data platform processing over one million records. The starting point was a 72% accurate model on the primary classification task - functional, but not production-ready. The goal was to improve accuracy while also reducing the preprocessing latency that was making the pipeline impractical to run at scale.

The accuracy improvement to 89% came from systematic feature engineering, not from swapping in a more powerful model. I ran structured ablations across feature sets, identified which signal sources were adding noise rather than signal, and redesigned the preprocessing pipeline so feature extraction was reproducible and auditable. The 40% latency reduction came from architectural changes to the data loading pipeline - specifically, eliminating redundant transformations that were being applied per-sample instead of once per batch.

This is the work that most ML engineering job descriptions describe but rarely see: not just achieving a metric, but understanding why the change worked and building infrastructure that makes the next improvement easier to measure. I led the 15-person cross-functional team as the senior ML engineer, responsible for architecture decisions, experiment tracking standards, and review of all model changes before they moved to evaluation.

At Auburn University from August 2023 to January 2024, I designed a reproducible ML evaluation framework that systematically benchmarked CNN architectures against classical methods across multiple medical imaging datasets. The framework made model selection decisions defensible - each architecture comparison was run under identical conditions with documented hyperparameter choices and reported with confidence intervals, not just point estimates.

I built experiment-tracking infrastructure for diabetic retinopathy classification that tracked every run: model architecture, training configuration, dataset split, and metric history. This infrastructure is why the findings contributed to a peer-reviewed publication - the experiments were reproducible by design, not because someone remembered to save a checkpoint.

At VIT CHAIR over two summers, I built end-to-end deep learning pipelines for Parkinson's detection and diabetic retinopathy classification, and signal-processing models using VMD and statistical features for automated sleep apnea classification. All work contributed directly to peer-reviewed publications. Shipping a pipeline that produces publication-quality results is a different engineering standard than shipping something that just runs - and that standard shapes how I approach ML evaluation in industry contexts too.