Model adaptation for expanding vocabularies
Examines how joint training degrades existing item and token embeddings while new entries are still learning.
Base Items Overfit, New Items Underfit: Hidden Cost of Joint Training in Incremental Adaptation
Senior Engineer, AI · Amazon Web Services
I develop AI systems at AWS. My earlier work includes model customization for Amazon Bedrock and recommendation systems for Amazon Personalize.
My current interests include post-training and inference at scale for foundation and custom models, as well as recommendation and search systems.
Publications
Amazon Nova foundation model reports
Selected work
Amazon Bedrock · Model customization and hosting
Primary architect for custom model deployment and technical lead for custom model import.
Amazon Personalize · Recommendation systems
Led development of new recommendation recipes that support larger item catalogs with lower inference latency.
Open source
- LangChain AWS: Bedrock chat integrations and AWS credential-based authentication.
- OpenSearch: Proposed a standalone OpenSearch integration package for LangChain; product quantization and reranking proposal referenced in the JVector integration discussion.
Professional service
- Area Chair for NeurIPS 2026 workshops: Verify-Agents and Managing Agents that Manage Agents (Meta-Agents); EMNLP 2026 workshop: GroundLM
- Program Committee member, IEEE International Conference on Agentic AI 2026
- Program Committee member, ICML 2026 workshops (CoLoRAI, DL4C); ACL 2026 workshop (ALVR); and ICLR 2026 workshops (SPOT, CAO, Agentic AI in the Wild)