I’m a PhD student at UCLA advised by Prof. Vwani Roychowdhury. My research centers on continual learning, memory, and reasoning in language models and agents: how these systems can learn from experience, adapt to changing tasks and environments, and ground their answers in evidence.
My work explores structured memory as a way for models to retain and use knowledge beyond a single context window. This includes episodic memory with Generative Semantic Workspaces (GSW) [6] [9], continual learning through external memory with Panini [5], and deep research with PANINI-DR [1], which brings together search, reasoning, and verification.
During my internships at AWS, I studied continual learning in enterprise agents, including how to evaluate adaptation when organizational rules change [3], and explored how non-parametric continual learning methods compare with tabular foundation models for forecasting [4].
I’ve also spent consecutive summers at NVIDIA contributing to Project G-Assist, an on-device AI assistant. My work spanned the entire stack, from tool-use data acquisition and fine-tuning to retrieval, evaluation, and inference optimization.
I’m also interested in AI for science and healthcare, with work on DNA sequence alignment [8], clinical information extraction [7], and clinical decision support [2]. In these settings, I’m particularly interested in uncertainty calibration—making confidence scores reflect how often predictions are correct—and using uncertainty to guide deferral to experts.
We introduce Panini-DR, a multi-agent system for deep research that uses Generative Semantic Workspaces to track subquestions, dependencies, findings, and supporting evidence. Separate research and verification agents check intermediate answers and guide further investigation across multi-step search and reasoning tasks.
We study LLM decision support for anti-seizure medication management in underrepresented epilepsy care, where systems must adapt to local prescribing practice and know when to defer. We introduce Manana, a multi-agent prompt-learning framework for site-specific prescribing corrections, and Bayesian prompt averaging, which treats learned prompt states as an ensemble to produce prescription probabilities, confidence scores, and deferral decisions for specialist review.
We introduce Continual Orchestration Bench to evaluate whether enterprise agents can learn from repeated workflows and adapt when organizational policies change. The benchmark tests orchestration across simulated services, exposing limitations in how current memory systems detect and recover from changing rules.
We study continual adaptation for forecasting by evolving a compact guidance document in a frozen LLM’s prompt. Outcome-driven reflection improves held-out forecasting and transfers across models, offering a training-free alternative to updating model weights.
We propose a continual learning framework where a fixed base model learns through an external semantic memory, representing documents as Generative Semantic Workspaces—networks of question-answer pairs that enable reasoning-grounded inference with superior performance while reducing token usage.
We introduce GSW, which equips LLMs with human-like episodic memory through structured semantic representations that track actors, roles, and states across space and time, outperforming standard RAG frameworks on episodic memory tasks.
We develop a framework using open-source LLMs to standardize opioid use disorder prescriptions from diverse EHRs, enabling consistent analysis across different EHR sources and formats.
We introduce DNARDE, a DNA language model trained with a contrastive objective for sequence alignment, pairing it with vector-store retrieval over nucleotide sequences. DNARDE outperforms prior transformer baselines on alignment and transfers across chromosomes and species.
Initial explorations on what an evolving memory system could look like from the perspective of how an ideal AI would observe and build a semantic map of an ongoing situation.
* denotes equal contribution.
NeurIPS LAW '25, AAAI '26, CAIS '26, ICML '26 Gold Reviewer, NeurIPS '26