Nov 2024 – Present · 1 yr 7 mos · Palo Alto, California, United States · On-site

• Working on Amazon Core Search Team
• Building Reasoning Models which are Search Aware
• Formulating scaling laws using model size and data composition for LLM Semantic Matching models (1B–20B scale)
• Spearheaded automated evaluation and query benchmarking to iteratively refine LLM Semantic Matching models using LLM as a Judge

May 2024 – Oct 2024 · 5 mos · San Francisco Bay Area · Remote

• Built general representation for solving various NLP problems on long financial documents
• (Built Financial Artificial General Intelligence)
• Lead a team of 3 for planning on training large foundation model for all NLP tasks across the company outlining relevant work, large compute scaling costs, data size estimations, model training debugging, and model inference optimization [Training 8B open-source variants such as Llama 3.1 8B, Mixtral-7B, and Gemma 1.1 7B]
• Engineered a document information retrieval system for the financial domain using a supervised fine-tuned Llama 3.1 8B model trained on a multi-GPU setup, resulting in a 55% improvement in analyst annotation efficiency
• Trained Llama 3.1 8B model with LoRA, Quantization and mixed precision using FSDP for multi-GPU training enabling data and model parallelism across 8 A100s (40 GB). Model results range in the order of 80% in exact match metrics on out of domain dataset for information extraction in JSON format.
• Explored the application of chain of thought in-context learning for information retrieval from long
• financial text using Claude 3.5 Sonnet showing 3% improvement compared to previous production models

Sep 2023 – May 2024 · 8 mos · San Francisco Bay Area · Remote

• Worked on applied NLP research in financial domain
• Built an end-to-end DeepSpeed-powered training framework for distributed multi-GPU training across all stages, from tokenization training, pre-training to downstream task fine-tuning, achieving a 40% improvement in model development efficiency
• Studied the application of in-context learning for Information retrieval from long financial documents using LLM showing 3% improvement as compared to current production models
• Deployed a multi-class text classification model based on sentence embeddings using AutoML improving upon the previous production model's performance by 8%
• Deployed an efficient longformer for long-context multi-class classification improving upon the previous production model’s performance by 10% on Kubernetes
• Orchestrated an offline calibration framework, and online regression testing framework to monitor performance, and rapidly perform iterative updates over production model

Jun 2023 – Dec 2024 · 1 yr 6 mos · Gainesville, Florida, United States

• Worked on a DARPA funded research project called Environment-driven Conceptual Learning (ECOLE) at the Data Science Research (DSR) lab UF. I am working under the guidance of Dr. Daisy Wang from UF, Dr. Eric Xing from CMU, and Dr. Zhiting Hu from UCSD.
• Worked on the alignment problem and multimodal reasoning
• Researched the application of reinforcement learning from human feedback for improving Knowledge Graphs and Scene Graphs
• Studied improvement, and faithfulness of multi-modal reasoning using in-context learning and human feedback for large language and vision models
• Investigated video reasoning with SOTA models like Qwen CogVLM2-Video (EVA-CLIP-E 12B + Llama 3 8B), and EMU2 (EVA-CLIP 4B + Llama 33B) to enhance multimodal comprehension.
• Examined methods to improve LLM reasoning faithfulness by refining chain-of-thought (CoT) processes through human and self-model feedback and subsequent training on corrected CoT

Dec 2022 – May 2023 · 5 mos · Gainesville, Florida, United States

• Worked on distance-sensitive range querying for high-dimensional data using deep learning at the Data Science Research (DSR) lab UF
• Orchestrated several experimental designs for training SelNet, Mixture of experts (MoE), XGB, LightGBM, and Support Vector Regressor on 3 text based and 3 image based embedding datasets using distributed training framework
• Composed the introduction, related work, and experimental details sections of the research paper for baseline experiments using Latex

Aug 2022 – Dec 2022 · 4 mos · Seattle, Washington, United States · On-site

• Worked on Alexa Smart Home Team
• Removed redundant features by performing feature importance analysis using visualization techniques, and model ablation study
• Adopted a novel self-attention based architecture for multi-variate time-series classification task to model user behaviors
• Performed mechanistic interpretability analysis on attention heads to uncover key insights from model layers

May 2022 – Aug 2022 · 3 mos · Seattle, Washington, United States · On-site

• Worked on Siri Text-to-Speech Team
• Design and implement an end-to-end acoustic model for text-to-speech synthesis
• Built a data synthesis and training pipeline to train deep learning models on large scale speech corpus (500 hours)

Mar 2022 – May 2022 · 2 mos · Gainesville, Florida, United States · On-site

• MISQA: A Dataset for Query-based Misinformation Mining
• Worked on developing a multi-answer QA Benchmarking Dataset at Data Science Research (DSR) lab UF
• Generated 1,000 QA pairs for the multi-answer QA benchmark dataset, to track misinformation and disinformation by prompting finetuned mT5 model
• Performed Tweet stance annotation for 1,000 samples QA pairs from Twitter API Querier using Dense Distinct Tweet Retriever (DDTR) model

Sep 2021 – Mar 2022 · 6 mos · Gainesville, Florida, United States · On-site

• Worked on Active Interpretation of Disparate Alternatives (AIDA) a DARPA Hypothesis Generation Project at Data Science Research (DSR) lab UF
• The project aims to automatically ingest web documents and transform them into a semantic space representation (Knowledge Graph) that analysts can use to query about uncertain situations and obtain a variety of related hypotheses
• Extracted sentence embeddings from 100,000 sentences using XLM Roberta for similarity clustering
• Improved data quality by performing text augmentation using prompt-based paraphrasing on Parrot model
• Built pipeline for cross-lingual Natural Language Inferencing using mT5 model on hand-annotated and prompt based paraphrased 5,000 cross-claim pairs, improving the score by 38%
• Developed and deployed a website to showcase the query-claims relationship using AngularJS and PrimeNG for 200 claims

Dec 2020 – Apr 2021 · 4 mos · Ahmedabad, Gujarat, India · Remote

• Retrieval of inherent optical properties (IOPs) of coastal waters using satellite data
• Worked on applied machine learning for oceanography using satellite data
• Researched and developed a modified Neural Network algorithm to solve the inversion problem of acquiring 6 IOPs from remote surface reflectance data.
• Achieved a good settlement with respect to R-Square values for each water quality parameter of about 97%.

May 2020 – Jul 2020 · 2 mos · Vadodara, Gujarat, India · Remote

• Worked on large-scale applications based on PHP, Laravel, and Javascript
• Refactored the codebase of 5 modules, reducing Npath and Cyclometric Complexity by 85%
• Improved search and sort speeds by 35% by fixing search and sort issues in 5 modules using custom AJAX calls
• Prepared a report on code complexity analysis on different modules using tools like code forensics
• Managed commit messages' length using commit message hooks in git.