Priyank Patel

Experience

1. Applied Scientist

   Amazon

   Apr 2021 — Present · Bengaluru, IN

   Inference latency and throughput optimization of transformer models in the Search team- Used inference latency estimates of transformer models derived from first principles to quantify relative latency reductions possible under a wide range of model compression and inference optimization techniques in order to determine the most promising combination of techniques for implementation and experimentation- Reduced the inference latency of encoder-only and encoder-decoder transformer models used for query understanding and query reformulation with a combination of weight pruning and post-training quantization techniques- Reduced the end-to-end text generation latency of a large language model (LLM) used for anaphoric and contextual dereferencing of customer questions in a multi-turn setting with a combination of prompt caching and compression, post-training quantization, and speculative decoding techniques that together reduced the LLM\'s time to first token as well as subsequent inter-token latencies- A part of this work was published at the Amazon Machine Learning Conference (AMLC) 2024.Developed machine learning models and algorithms for various problems in the Search team- Identifying vertical search intent in queries to improve customer search experience while working with limited, sparse, and imbalanced behavioral data for training transformer based classification models- Expanding coverage of numeric behavioral features used in search relevance ranking with transformer based regression models- Machine learnt generic keywords to improve discoverability of new products in search.

Education

• The University of Texas at Austin

  Master of Science - MS, Electrical Engineering

  2005 — 2007

• Illinois Institute of Technology

  Bachelor of Science - BS, Electrical Engineering

  2001 — 2005

• The University of Texas at Austin

  Further Graduate Studies, Electrical Engineering

  2008 — 2012