Smarak K.

Key Skills

Freelancer Overview

I have extensive hands-on experience orchestrating and curating massive, unstructured datasets to train and evaluate high-performance AI systems. In the domain of Generative AI and LLMs, I processed and structured Terabytes of video dataset files to probe 3D scene understanding and novel view synthesis in my work at the University of Toronto. I have also curated a Multilingual MATH-500 dataset with cross-lingual paired splits to isolate token-level reasoning failures. My work with large-scale scientific data includes managing a massive planetary dataset of over 7 million individual files containing raw XRF Spectroscopy data from ISRO’s Chandrayaan-2 mission, where I built parallelized pipelines across a distributed system of 10+ servers to preprocess and align the unstructured spatial readouts for high-resolution elemental mapping. Complementing this, I have deep experience engineering datasets for complex geospatial and multimodal analytics. During my work with large-scale satellite imagery and demographic surveys at Stanford University, I successfully parsed through highly unstructured data containing 11,945 variables, engineering a feature-reduction strategy that cut dataset dimensionality by 80% while retaining the high predictive integrity needed for downstream deep learning models. Additionally, during my time at Amazon, I worked on adapting pipelines for contextual data extraction from diverse product images, ensuring dataset robustness across varying visual contrasts and brightness levels. Across all these projects, my focus has been on transforming raw, messy, and high-dimensional data into cleanly structured, production-ready assets designed for rigorous model training and evaluation.