Devang Kulshreshtha
I’m a Senior Research Engineer on the Embodied AI team at Cruise and a final-year ML PhD student at NYU. I believe AI safety must shape how models are trained, understood, and deployed. My research spans AI safety [2, 4], agentic safety [3, 5], model adaptation [6, 9], and vision–language models [1]. I currently train foundation perception models for safer autonomous driving.
Previously, I led the prompt-attack filter for Bedrock Guardrails at AWS as an Applied Scientist, translating adversarial-robustness research into production defenses against jailbreaks and prompt injection. I also contributed to AgentCore Policy to govern agents’ access to tools. My goal is to make increasingly capable AI systems safe enough to trust in the real world.
Before my PhD, I completed my master’s in CS at McGill/Mila and my bachelor’s at IIT (BHU) Varanasi.
Experience
-
Senior Research EngineerSep 2026–presentPerception models for safer autonomous driving.
-
AmazonApplied Scientist IIOct 2023–Sep 2026AgentCore Policy for tool-use governance; led Bedrock Guardrails’ prompt-attack filter.Applied Scientist IJul 2022–Sep 2023Retrieval-augmented speech recognition and contextual adaptation.Applied Science InternMay–Sep 2021Language-model rescoring for speech recognition.Machine Learning Engineer2018–2020Anomaly detection for truck driver invoices.
-
Korbit.AIResearch Engineer2020–2022Neural question generation, retrieval, and adaptive feedback.
-
INRIAResearch InternSummer 2018Subgraph mining to interpret neural network failures.
Education
- New York UniversityOngoingPhD in ECE (ML) · GPA 4.0 / 4.0
- McGill University & Mila2020–2022MSc in Computer Science · GPA 4.0 / 4.0
- IIT (BHU) Varanasi2014–2018B.Tech in Computer Science · GPA 9.11 / 10
Publications & preprints
-
Multimodal View Training: Learning More from the Same Information
-
Break Me If You Can: Self-Jailbreaking of Aligned LLMs via Lexical Insertion Prompting
-
The Subtle Art of Defection: Understanding Uncooperative Behaviors in LLM based Multi-Agent Systems
-
STAC: When Innocent Tools Form Dangerous Chains for LLM Agents
-
Relational Priors as Convergence Pressure in LLM-Based Multi-Agent Systems
-
Sequential Editing for Lifelong Training of Speech Recognition Models
-
Retrieve and Copy: Scaling ASR Personalization to Large Catalogs
-
Generalized Zero-Shot Audio-to-Intent Classification
-
Multilingual Contextual Adapters for Custom Word Recognition in Low-Resource Languages
-
Improving Domain-Adaptive Generalization of CTC-Based ASR with Internal LM Estimation
-
Few-Shot Question Generation for Personalized Feedback in Intelligent Tutoring Systems
-
Back-Training Excels Self-Training at Unsupervised Domain Adaptation of Question Generation and Passage Retrieval
-
How Emotional Are You? Neural Architectures for Emotion Intensity Prediction in Microblogs
-
NLPRL-IITBHU at SemEval-2018 Task 3: Combining Linguistic Features and Emoji Pre-trained CNN for Irony Detection
-
Feature Augmented Deep Neural Networks for Collaborative Filtering
-
Prayas at EmoInt 2017: An Ensemble of Deep Neural Architectures for Emotion Intensity Prediction in Tweets
Patents
- Augmenting automated speech recognition neural networks with scalable vocabularies. [US 12562151B1]
- Removing bias from speech recognition models using internal language model estimates. [US 12387718B1]
- Automated evaluation of evidence mapping models. [US 20250005063A1]
- Medical conversation summarization style intelligence. [US 20240428002A1]
Service & recognition
Reviewing
COLM 2026 · ACL 2026 · AAAI 2026
INTERSPEECH 2024, 2025
ACL Rolling Review 2022, 2024
Teaching
IVADO Workshop, McGill / Mila
Introduction to AI, IIT Varanasi
Introduction to Programming in C, IIT Varanasi
Recognition
MITACS Accelerate Research Scholarship
Kharusi International Science Fellowship
Shared-task winner, WASSA @ EMNLP 2017