Devang Kulshreshtha

AI Researcher

United States5 yrs experience

Key Highlights

Led development of AWS Bedrock Guardrails Prompt Attack solution.
Contributed to major launches like HealthScribe and Lex.
Published 5+ papers and secured 6 patents in AI.

Stackforce AI infers this person is a leading expert in AI and machine learning applications.

Contact

Skills

Core Skills

Machine LearningNatural Language Processing (nlp)

Other Skills

Large Language Models (LLM)Speech Recognition

About

Hi! I’m Devang, an Applied Scientist II at Amazon AWS Agentic AI in New York, with 5+ years of experience in machine learning research and engineering. My research interests include safety and red-teaming in LLMs, Agentic AI Safety, Speech Recognition (ASR), and NLP. At AWS, I led the development of the Bedrock Guardrails Prompt Attack defense, and contributed core science to launches such as HealthScribe and Lex, advancing summarization and ASR systems at scale. My work has resulted in 5+ publications (EMNLP, INTERSPEECH, IJCAI) and 6 patents spanning AI safety, ASR personalization, and lifelong learning. I completed my MSc in Computer Science at McGill University & Mila Lab under Prof. Siva Reddy, where I researched question–answer generation and personalised feedback systems with Korbit.AI.

Experience

5 yrs

Total Experience

1 yr 9 mos

Average Tenure

Current Experience

Amazon web services (aws)

2 roles

Applied Scientist II

Promoted

Oct 2023 – Present · 2 yrs 6 mos · On-site

Engineered the AWS Bedrock Prompt Attack solution, beating Microsoft and Nvidia solutions by 35%.
Inference Optimisation of AWS Guardrails Denied Topic Multilingual model for long-context inputs.
Developed lifelong training methods for ASR models to improve performance without full retraining.

Large Language Models (LLM)Machine LearningNatural Language Processing (NLP)Speech Recognition

Applied Scientist

Jul 2022 – Oct 2023 · 1 yr 3 mos · On-site

Building automatic speech recognition (ASR) models for low-resource languages, and scaling ASR personalisation to very large catalogs (>500K size).

Large Language Models (LLM)Machine LearningNatural Language Processing (NLP)Speech Recognition

Amazon

2 roles

Applied Scientist

Promoted

Jul 2021 – Aug 2021 · 1 mo · Cambridge, England, United Kingdom · On-site

Improved long-tail performance of ASR by 12% by developing robust LLM rescoring systems.
Secured a return full-time offer based on demonstrated research and engineering impact.

Large Language Models (LLM)Machine LearningNatural Language Processing (NLP)Speech Recognition

Software Engineer

Sep 2018 – Sep 2020 · 2 yrs · New Delhi Area, India

Designed a scalable anomaly detection system for invoice monitoring, cutting false positives by 30%.
Mentored junior developers and provided critical service support for high-availability systems.

Mila - quebec artificial intelligence institute

Graduate Student

Sep 2020 – Oct 2022 · 2 yrs 1 mo · Montreal, Quebec, Canada · On-site

Neural question generation from open educational resources like Wikipedia, specifically in the statistics and Machine Learning domains. Published 2 papers in top-tier ML conferences (EMNLP, IJCAI)

Large Language Models (LLM)Machine LearningNatural Language Processing (NLP)Speech Recognition

Korbit ai

Artificial Intelligence Researcher

Sep 2020 – Jun 2022 · 1 yr 9 mos · Montreal, Quebec, Canada · Hybrid

Built ML systems for automatic question-answer generation from educational materials
Developed NLP-based personalised feedback generation for intelligent tutoring systems.

Large Language Models (LLM)Machine LearningNatural Language Processing (NLP)Speech Recognition

Inria

Summer Research Intern

May 2018 – Aug 2018 · 3 mos · Rennes, Brittany, France

Mining Activation Patterns in Deep Neural Networks to identify neuron sets responsible for making incorrect predictions

Large Language Models (LLM)Machine LearningNatural Language Processing (NLP)Speech Recognition

Amazon

Software Engineer Intern

May 2017 – Jul 2017 · 2 mos · New Delhi Area, India

Designed and developed the "Amazon Trucker Android App" that lets owneroperators onboard and allocate drivers during the onboarding process.
Implemented rich material design on react-native framework and Amazon DynamoDB as the backend database service.

Busigence

Associate Researcher

Dec 2016 – Jan 2017 · 1 mo · Gurgaon, India

The project focused on constructing Deep learning frameworks for Recommender Systems.
Coded RBMs, Denoising AutoEncoders and other architectures from scratch for the collaborative filtering task.
Modified RBM to transform item attributes into domain-independent latent features. Item latent features alongside user transaction history and demographics are used to construct user latent features.