Projects
Tracked but Suppressed: How LLMs Fail Current-Value Retrieval
Interactive research showcaseAn interactive look at proactive interference in LLMs, extending our arXiv preprint with mechanistic analysis and a fix — feel the failure firsthand, see it hold across 19+ models, look inside the model at the mechanism, and see the small fine-tune that closes the gap. A dedicated paper on the extended results is in preparation.
Programming with Representations (PwR)
Microsoft Research IndiaCreated an intermediate representation layer to bridge users and Large Language Models, addressing hallucination and enabling non-technical users to build AI-powered applications.
Multimodal Ingredient Substitution Knowledge Graph
Master's Thesis · AI Institute, USCBuilt a knowledge graph incorporating 27K ingredients and 40K substitution pairs, enabling precise ingredient recommendations using multimodal and constraint-based search. Developed an LLM-based query module and contributed to AAAI-25 submission.
Cook-Gen
IEEE SMC 2023Formulated cross-modal recipe retrieval and developed cooking action recognition for recipe analysis, achieving 95% recall. Published at IEEE SMC 2023.