Postdoctoral Researcher - The Conservation Fund Freshwater Institute - Shepherdstown, WV
(2024-08 - 2026-08)
- Computer vision pipeline, prototype to production: Led end-to-end development of a hyperspectral computer-vision pipeline (Python/OpenCV) for biological classification and automated ROI segmentation, achieving 81.8% classification accuracy and 98.8% mAP@50 on real-world production data.
- Edge and cloud model benchmarking: Designed and executed a systematic benchmark of 12 object detection model variants (YOLOv5u, YOLOv8, YOLO11, YOLO26) across 7 training dataset sizes, evaluating accuracy, data efficiency, and inference latency on GPU (A100) and edge (Raspberry Pi 5) hardware.
- Model validation and robustness: Designed a leave-one-date-out validation protocol that exposed session-specific distribution shift and confounding invisible to standard cross-validation, directly testing whether a model's predictions held up outside their training conditions.
- Production deployment: Contributed to prototyping FilletCam 2.0 from early development, including hardware assembly, a YOLOv8-based computer vision tool for automated quality and defect assessment, now commercially deployed and in active use by industry customers.
- Applied geospatial ML: Built a Google Earth Engine and Python monitoring workflow fusing satellite imagery with in-situ sensor data across 20 sites, delivering a decision-support tool for non-technical stakeholders.
Graduate Research Assistant (PhD Dissertation) - Washington State University - Pullman, WA
(2019-08 - 2024-07)
- Sensor-to-image modeling: Applied PyTorch to reformulate high-dimensional sensor spectra as 2D image data for CNN-based classification of plant disease biomarkers, validated against molecular (qPCR) ground truth at 88% accuracy.
- Statistical model benchmarking: Benchmarked 9 classifiers (Random Forest, SVM, Logistic Regression, Gradient Boosting, LDA/QDA, k-NN, Naive Bayes) via PCA and 5-fold cross-validation, with PLS-VIP, ANOVA, and Tukey HSD statistical analysis.
- Geometric and 3D signal processing: Built UAV-derived 3D point-cloud pipelines (Python, Open3D: convex hull, alpha-shape, voxel-grid algorithms) for canopy segmentation and volume estimation from multi-view field imagery, published first-author at IEEE MetroAgriFor 2021.
- Field data collection and calibration: Planned and flew multispectral and thermal UAV imaging missions as part of a DOE/PNNL project, radiometrically calibrated imagery against reflectance panels, and used RTK ground control for georeferencing across repeated acquisitions.
- Quantified systematic bias between satellite and UAV vegetation-index measurements across three growing seasons, a data-quality finding that changed how the team interpreted downstream results.
- Teaching Assistant for two graduate courses, Instrumentation and Measurements (BSYSE 541) and Unmanned Aerial Systems in Agriculture (BSYSE 552).
- Mentored undergraduate interns and junior lab members on experimental methods and data analysis.
Data Scientist, Geospatial ML - Quantela Inc. - Bangalore, India
(2017-06 - 2019-07)
- Decision-support dashboards: Built land-use classification and flood-risk mapping models from large-scale satellite imagery, translating raw sensor data into risk assessments and dashboards that public-sector stakeholders used directly for decision-making.
- Data infrastructure: Architected AWS-based pipelines for large-scale satellite data ingestion and feature engineering, building a reproducible methodology applied consistently across new regions and datasets.
Research Intern - Indian Institute of Tropical Meteorology (IITM) - Pune, India
(2017-01 - 2017-06)
- Atmospheric spectral analysis: Analyzed 26 years (1979-2004) of ERA reanalysis and LMDZ4 climate model data to characterize atmospheric variability in the Karakoram and Central Himalaya, isolating synoptic-scale signals via Lanczos high-pass filtering on HPC.
- Periodicity discovery: Applied power spectral density analysis to the filtered data, identifying a dominant 3 to 8 day periodicity in Western Disturbance activity, consistent across both reanalysis and model datasets.
Earlier Research Roles - TCS, KisanHub, IIRS/ISRO, India - India
(2013 - 2016)
- Mapped spatio-temporal rice cultivation across 330,000+ hectares using Landsat 8 spectral indices and ensemble ML, achieving 87.99% accuracy for land-use monitoring (TCS Innovation Labs).
- Automated FAO-56 evapotranspiration estimation for 500+ farm parcels and built satellite-data processing workflows with unit testing and QA (KisanHub).
- Mapped land degradation (waterlogging, soil salinity) across 1,868 km² in Haryana using Landsat-8 and GIS, validating results against lab-tested soil pH, electrical conductivity, and moisture content (IIRS/ISRO).
