May 2020 – Aug 2020 · 3 mos · Austin, Texas, United States · Remote

• Data & AI Services (DAIS) Team in ARM Research
• Abstracted the problem of advertising recommendations based on user clicks and developed a variational autoencoder to encode ad slates and reinforcement learning-based approach to improve funneling.

Jun 2019 – Aug 2019 · 2 mos · Austin, Texas Metropolitan Area

• Machine Learning Team
• Developed an AI risk assessment tool to score black-box regression models on 4 metrics, including explainability and fairness. This has been productionized in Cortex CERTIFAI.
• Interacted with industry stakeholders and proposed a responsible AI framework for insurance. Team placed first in company-wide Shark Tank competition on products for responsible AI.

Jan 2019 – May 2019 · 4 mos

• Teaching assistant for EE 351M - Digital Signal Processing
• Held regular office hours for over 60 students to answer questions pertaining to the course material and homeworks.

May 2018 – Aug 2024 · 6 yrs 3 mos

• Haplotype Assembly using Long-distance Read Correlations Learned by Transformer
• Pioneered a transformer-based approach, XHap, for reference-based haplotype assembly from sequencing reads. XHap infers a 40% richer correlation structure and reconstructs > 300× longer haplotypes accurately.
• Incorporated parallelization across 4 GPUs and preprocessed data using Bash and C++ scripts for benchmarking XHap on > 100GB of human sequencing reads.
• Transformer-based Oncoviral Read Identification from Cancer Samples
• Led the development of a pipeline in PyTorch (XVir) for oncoviral read identification through the use of a transformer-based classifier, achieving 99% accuracy.
• XVir trains > 8× faster than SOTA methods and robust to as much as 15% diversity in the oncoviral families.
• Differentially Private Median Forests
• Proposed a median-based approach to construct differentially private decision trees to privacy preserving noise in the splits and node statistics.
• The superior accuracy of the designed approach was validated with error bars over 10 datasets for both regression and classification and 10+ hyperparameter settings.
• Reconstructing Intra-tumor Heterogeneity via Convex Optimization and Branch-and-Bound Search
• Formulated the problem of copy number inference in tumors as matrix factorization.
• Designed computationally efficient and accurate approach in MATLAB that sped up inference almost 10×.

Aug 2017 – May 2018 · 9 mos

• Teaching assistant for EE 362K - Introduction to Automatic Control
• Held regular office hours for over 60 students to answer questions pertaining to the course material and homeworks.
• Helped in formulating homeworks and examinations.

May 2016 – Jul 2016 · 2 mos · Cambridge, Massachusetts

• Developed a set-up on a development board to measure power consumption of on-board CPU and GPU of any code. This could then be used with a profiler to identify areas for optimization.

May 2015 – Jul 2015 · 2 mos · Greater Oxford Area