hello!

LLMs are increasingly employed as surrogate annotators and evaluators in NLP workflows. However, current practices often involve multiple heuristic decisions to design effective workflows. For example, choosing the appropriate subset to annotate — by LLMs or humans — remains a costly decision.

Recently, I've been working on AnnotationArena — an end-to-end framework to streamline LLM-based evaluation and annotation. This includes:

• Using Value of Information [1], [2] for inference-time decision making.
• Leveraging gradient-based heuristics [3], [4] for active learning.
• Exploring reinforcement learning and alignment techniques to enable adaptive, continuous annotation pipelines with principled decisions.

I'm also interested in alternative labeling strategies such as ratings, rankings, and ordinal classifications.

I build models that integrate speech, text, and vision for tasks like:

• Emotion Recognition (example)
• Speaker Diarization (example)

My models learn from heterogeneous modalities with minimal supervision. I began my research journey with emotion recognition, and while I’ve developed a fair amount of expertise in that space, I’m increasingly drawn to speaker recognition and diarization. I find diarization particularly interesting — it's a fundamental speech task with open challenges in temporal structure, multimodal fusion, and low-resource adaptation. My recent works are focused on improving diarization quality through multitask learning and adaptive fusion.

Thanks to rigorous theory coursework at JHU (this), I’ve developed a strong interest in replicability theory — distinct from reproducibility (See [this] and [this]).

My initial research in this area (guided by Prof. Jess Sorrell) focuses on:

• Deriving replicability bounds for transfer learning.
• Investigating how adaptive data selection affects transferability and stability of learned models.

Read our ongoing manuscript: Sensitivity of Selectivity in Transfer Learning (work-in-progress).