Resume
Experience
Applied Scientist — Ambient Perceptual Technologies (Amazon AGI)
June 2022 – Present · Sunnyvale, CA
- Developed and productionized a unified “world encoder” for multi-task training across 17 Alexa locales (KWS, LV ASR, DVAD). Architected locale- and task-specific decoders on a shared encoder backbone, reducing cross-locale model duplication while achieving production-level parity. Deployed to scale in Spanish locales, currently undergoing live A/B testing before ramping across 6 additional locales.
- Built and productionized on-device Device Voice Activity Detection (DVAD) models to determine whether speech was directed to Alexa devices. Served as the gating component for the Alexa+ en_US launch, operating in production at scale.
- Developed production-grade keyword spotting models for automotive applications with limited in-domain data and challenging acoustic conditions, achieving 56% lower FRR and 71% lower FA on enhanced Siri certification tests compared to previous baselines.
- Designed and deployed a QAT-enabled model architecture across 10 Alexa locales, improving FRR by ~8.5% in offline evaluations and online recall by 4–16%, depending on locale.
- Improved fairness across gender-based cohorts, reducing FRR for feminine-sounding voices by 10.1% and for masculine-sounding voices by 7.8%, resulting in a 27% relative reduction in FRR gap across 16 locales.
- Implemented Quantization-Aware Training (QAT) and Knowledge Distillation (KD)-based model compression methods, demonstrating that sub-8-bit models maintain performance on par with full-precision variants.
- Developed and launched ML models for Amazon Greenwood auto devices in UK and Spain, achieving ~41% and ~64% FRR improvements respectively, enabling product launch in EU markets.
Software Development Engineer — Amazon Web Services (AWS)
June 2021 – June 2022 · Santa Clara, CA
- Designed and implemented a scalable metrics framework to evaluate annotation quality across multiple modalities, improving dataset reliability and downstream model performance.
- Re-architected the internal labeling workflow to dynamically partition workloads by label category, mitigating class imbalance and reducing annotator cognitive load.
Software Development Engineering Intern — Amazon (Alexa)
May 2020 – July 2020 · Seattle, WA
- Designed and implemented a feature within Alexa Smart Home (SHCC), enabling spatial visualization of device groupings by overlaying contextual metadata onto home layout topology.
- Contributed to the Context and Targeting (Spaces) system, improving customers’ ability to manage and visualize smart home device groups.
Graduate Student Researcher — Purdue University
August 2020 – May 2021 · West Lafayette, IN
- Proposed a novel personalization framework for Federated Learning by unifying knowledge distillation with user-specific optimal teacher models, improving client-level performance in heterogeneous data settings.
- Introduced new metrics to evaluate performance and fairness in personalized FL systems beyond average accuracy. Work presented at the ICML 2021 Federated Learning Workshop and published on arXiv.
Education
Purdue University — West Lafayette, IN
M.S. in Computer Science (Machine Learning, Data Science — Thesis Track) · GPA: 4.0 · May 2021
SSN College of Engineering — Chennai, India
B.E. in Computer Science and Engineering · CGPA: 8.79/10 · May 2019
Publications
New Metrics to Evaluate the Performance and Fairness of Personalized Federated Learning Poster at the Federated Learning Workshop, ICML 2021. Also on arXiv. (June 2021)
Unifying Distillation with Personalization in Federated Learning arXiv. (May 2021)
DeepTrace: A Generic Framework for Time Series Forecasting International Work-Conference on Artificial Neural Networks (IWANN), Gran Canaria, Spain. (June 2019)
Forecasting Food Sales in a Multiplex Using Dynamic Artificial Neural Networks Computer Vision Conference (CVC), Las Vegas, USA. (April 2019)
Convolutional Long Short-Term Memory Neural Networks for Hierarchical Species Prediction Conference and Labs of the Evaluation Forum (CLEF), Avignon, France. (September 2018)