AI Safety · LLM Research · Agentic AI

Devang
Kulshreshtha

Applied Scientist II · Amazon AWS  |  PhD · NYU

Researcher and applied scientist with 5+ years in machine learning research and engineering, focused on LLM safety and red-teaming, post-training (SFT, RLHF, model editing), and agentic AI systems. Previously worked on speech recognition, multilingual NLP, and continual learning.

Google Scholar arXiv LinkedIn GitHub
Devang Kulshreshtha
🎓
PhD in Machine Learning (part-time)
New York University
2025 – Present
🎓
MSc, Computer Science  GPA 4.0/4.0
McGill University & Mila Lab
2020 – 2022 · Advisor: Siva Reddy
🎓
B.Tech, Computer Science  9.11/10
IIT Varanasi (BHU)
2014 – 2018
Career

Experience

Amazon AWS AI  (AgentCore · Bedrock · Lex)
Applied Scientist II
Oct 2023 – Present · New York
  • Research on LLM safety and red-teaming, developing novel attack and defense techniques for production guardrail systems.
  • Design and deployment of ML safety classifiers for agentic AI workflows.
  • LLM post-training: domain-adaptive pre-training, SFT, and RLHF for policy enforcement in agentic systems.
  • Synthetic data infrastructure design for cold-start agentic AI scenarios.
  • Multimodal LLM evaluation framework spanning speech, language, and audio benchmarks.
  • Mentored research interns to top-venue publications; reviewer for ACL, EMNLP, AAAI, INTERSPEECH.
Amazon AWS AI  (Bedrock · Lex)
Applied Scientist I
Jun 2022 – Sep 2023 · Seattle
  • Research on retrieval-augmented methods for scaling ASR personalization to large entity catalogs. Deployed to production. (EMNLP 2023)
  • Multilingual contextual adapters for custom word recognition, expanding production ASR coverage across languages and locales. (INTERSPEECH 2023)
Korbit.AI
Research Engineer
2020 – 2022 · Montreal
  • Transformer and retrieval-based models for automated question generation and adaptive feedback in large-scale intelligent tutoring systems. (EMNLP 2021, IJCAI-PAIS 2022)
Amazon AWS  (Internship)
Applied Science Intern
May – Aug 2021 · UK (Remote)
  • Research on LLM-based rescoring for improving ASR accuracy on long-tail European languages.
Amazon India
Software Engineer
2018 – 2020 · Bangalore
  • Built scalable anomaly detection systems for invoice monitoring, reducing false positives by 30%.
INRIA Research Labs · University of Rennes
Research Intern
Summer 2018 · France
  • Subgraph mining approach for identifying DNN computation patterns linked to incorrect predictions, improving diagnostic accuracy.
Research Output

Publications

2026
Break Me If You Can: Self-Jailbreaking of Aligned LLMs via Lexical Insertion Prompting
D. Kulshreshtha, H. Su, H. Wang, C. Hegde
TrustNLP @ ACL 2026 Paper
2026
Understanding Uncooperative Behaviors in LLM-based Multi-Agent Systems
D. Kulshreshtha, W. Du, R. Jain, S. Doss, H. Su, S. Swamy, Y. Qi
EACL 2026 (Industry) Paper
2026
STAC: When Innocent Tools Form Dangerous Chains to Jailbreak LLM Agents
J.J. Li, J. He, D. Kulshreshtha, X. Xian, Y. Zhang, H. Su, S. Swamy, Y. Qi
arXiv 2026 Paper
2024
Sequential Editing for Lifelong Training of Speech Recognition Models
D. Kulshreshtha, S. Dingliwal, B. Houston, N. Pappas, S. Ronanki
INTERSPEECH 2024 Paper
2023
Retrieve and Copy: Scaling ASR Personalization to Large Catalogs
S.M. Jayanthi, D. Kulshreshtha, S. Dingliwal, S. Ronanki, S. Bodapati
EMNLP 2023 (Industry) Paper
2023
Generalized Zero-Shot Audio-to-Intent Classification
V.R. Elluru, D. Kulshreshtha, R. Paturi, S. Bodapati, S. Ronanki
IEEE ASRU 2023 Paper
2023
Multilingual Contextual Adapters for Custom Word Recognition in Low-Resource Languages
D. Kulshreshtha, S. Dingliwal, B. Houston, S. Bodapati
INTERSPEECH 2023 Paper
2023
Improving Domain-Adaptive Generalization of CTC-Based ASR with Internal LM Estimation
N. Das, M. Sunkara, S. Bodapati, J. Cai, D. Kulshreshtha, J. Farris, K. Kirchhoff
ICASSP 2023 Paper
2022
Few-Shot Question Generation for Personalized Feedback in Intelligent Tutoring Systems
D. Kulshreshtha, M. Shayan, R. Belfer, S. Reddy, I.V. Serban, E. Kochmar
IJCAI-PAIS 2022 Paper
2021
Back-Training Excels Self-Training at Unsupervised Domain Adaptation of Question Generation and Passage Retrieval
D. Kulshreshtha, R. Belfer, I.V. Serban, S. Reddy
EMNLP 2021 Paper Code
2018
How Emotional Are You? Neural Architectures for Emotion Intensity Prediction in Microblogs
D. Kulshreshtha, P. Goel, A.K. Singh
COLING 2018 Paper Code
2018
NLPRL-IITBHU at SemEval-2018 Task 3: Combining Linguistic Features and Emoji Pre-trained CNN for Irony Detection
H. Rangwani, D. Kulshreshtha, A.K. Singh
SemEval @ NAACL 2018 Paper
2017
Feature Augmented Deep Neural Networks for Collaborative Filtering
D. Kulshreshtha
IJCAI 2017 Workshop Paper Code
2017
Prayas at EmoInt 2017: An Ensemble of Deep Neural Architectures for Emotion Intensity Prediction in Tweets
P. Goel, D. Kulshreshtha, P. Jain, K.K. Shukla
WASSA @ EMNLP 2017 Paper
Intellectual Property

US Patents

US 12387718B1
Removing Bias from Speech Recognition Models using Internal Language Model Estimates
N. Das, M.N. Sunkara, S.B. Bodapati, J. Cai, D. Kulshreshtha, J. Farris, N. Aldridge
Granted  View
US 20250005063A1
Automated Evaluation of Evidence Mapping Models
D. Kulshreshtha, S. Dingliwal, S.B. Bodapati, K. Kirchhoff, S. Handa
Granted  View
US 20240428002A1
Medical Conversation Summarization Style Intelligence
A. Elangovan, D. Kulshreshtha, K. Kirchhoff, L. Xu, S. Handa, S.B. Bodapati
Granted  View
US 12562151B1
Augmenting Automated Speech Recognition Neural Networks with Scalable Vocabularies
D. Kulshreshtha, S. Dingliwal, S.B. Bodapati, V.R. Elluru, A. Mishra, K. Kirchhoff
Granted  View
Filed
Sequential Editing for Lifelong Training of Speech Recognition Models
D. Kulshreshtha, N. Pappas, S. Dingliwal, B. Houston, S. Ronanki, V.R. Elluru
Filed
Filed
Multilingual Contextual Adapters for ASR Personalization of Low-Resource Languages
D. Kulshreshtha, S. Dingliwal, B. Houston, S. Bodapati, S. Ronanki, et al.
Filed
Community

Academic Service

Reviewing

  • COLM 2026
  • ACL 2026
  • AAAI 2026 — Main Technical Track
  • INTERSPEECH 2025
  • ACL Rolling Review 2024
  • INTERSPEECH 2024
  • ACL Rolling Review 2022

Talks

  • INTERSPEECH 2023 — Dublin, Ireland
  • IJCAI-PAIS 2022 — Vienna (Virtual)
  • EMNLP 2021 — Virtual
  • COLING 2018 — Virtual
  • IJCAI Workshop 2017 — Melbourne

Teaching

  • McGill University / Mila — IVADO Workshop
  • IIT Varanasi — Introduction to AI (Spring 2018)
  • IIT Varanasi — Intro to Programming in C (Fall 2016)
Get in Touch

Let's Connect

Interested in discussing AI safety research, LLM post-training, or agentic systems? Reach out.

Email LinkedIn Google Scholar GitHub